MentalChat16K: A Benchmark Dataset for Conversational Mental Health Assistance

W4.2 The reviewer appreciates the detailed explanation of each score to be used in the prompt for LLM-based evaluation, but since this is such a detailed and nuanced range of scores, would the LLM really be capable of differentiating between different scores? Showing a distribution of LLM-based evaluation score one or few factors would be helpful. For example, it would point to potential issues if the resulting scores across 200 examples concentrate around "typical good" values, like 7 or 8.

Thank you for raising this insightful question. To address your concern about whether the LLM is capable of distinguishing nuanced scores across examples, we analyzed the distribution of scores assigned by two LLM judges, Gemini-Pro and GPT-4 Turbo, across 200 evaluation examples for all 7 metrics. The table below presents the frequency of each score for the two LLM judges. Please see Appendix A.8 for visualization of this result.

The results show that while the scores tend to cluster around 7 to 9—reflecting the generally high quality of responses—the score distributions also demonstrate variability across the full range (1–10). For higher scores, both histograms indicate that the LLMs are not treating all high-quality responses as equivalent. Instead, they appear capable of distinguishing between varying levels of quality within the "good" range, such as differentiating between a solid response (7) and an exceptional one (9). For lower scores, the GPT-4 Turbo evaluation demonstrates greater variability, with a more evident presence of lower scores (e.g., 1–6). This indicates both LLM’s ability to identify and differentiate between responses of varying quality.

W4.3 Are the seven-factor metrics employed only in LLM-based evaluation, or do human experts follow along the same line of logic when evaluating the quality of the response?

Yes, the seven-factor metrics are employed only in LLM-based evaluation and we used rank as the metric for human evaluation. The reason is that it is difficult to let humans distinguish such nuanced differences between the seven metrics. It is also easier for humans to have an overall impression on each response to the same question, which reduces the evaluation burden while preserving the accuracy of the evaluation.

W4.4 Since human evaluation is involved, and multiple experts have worked on the same question during evaluation, it would be more robust to include inter-rater agreement between the experts (or at least the variance in human ratings in addition to simple mean).

Thank you for raising this important point. We agree that including inter-rater agreement or the variance in human ratings would strengthen the robustness of our evaluation. To address this concern, we calculated the inter-rater reliability using Cohen’s kappa score. The average kappa score across all pairs of evaluators was 0.441, which falls into the “moderate agreement” range as per the Landis and Koch (1977) scale [1]. This score reflects a reasonable level of consistency among the evaluators while also highlighting opportunities for refinement. For further details on the guidelines provided to human evaluators, the specific criteria used for ranking, and steps taken to ensure consistency, we kindly refer you to Section A1 of the General Response 2/3 above. The details are also discussed in Section 3.3 from Lines 312 to 323 in the revised manuscript.

We appreciate your feedback, which has helped us recognize the need for even greater rigor in our evaluation protocols. In future studies, we will aim to improve inter-rater agreement through enhanced evaluator training, clearer rubrics, and additional reliability metrics.

[1] Landis, J. R., & Koch, G. G. (1977). The measurement of observer agreement for categorical data. Biometrics, 33(1), 159-174.

Continued from Response to A2qn 4/6:

W6.2 Conduct similarity-based analysis between questions in the synthetic data, in the interview data, and in the evaluation set. Reddit is a public data source that could be used in the training of GPT models, so this could reveal potential data leakage issue in the synthetic data. This is not an invalidation of the synthetic data, but more revealing the lack of efficacy in the evaluation approach. And it could also help reveal if there's an obvious domain shift between interview data and synthetic data.

To address your concern about potential data leakage and domain shift, we conducted a similarity analysis using Sentence-BERT to calculate cosine similarities between questions in our synthetic data, interview data, and evaluation set. We randomly selected 10 depression-related questions from each dataset and found the following average similarities:

• Synthetic vs. Interview: 0.3490
• Synthetic vs. Evaluation: 0.2829
• Interview vs. Evaluation: 0.3338.

These moderate similarity scores indicate that while there is some overlap across datasets, they are not excessively similar, which reduces the risk of data leakage from Reddit into our synthetic data. The higher similarity between the synthetic and interview data suggests that our synthetic data effectively captures the nuances of real interview questions without duplicating them. Additionally, the comparable similarities across datasets imply there is no significant domain shift, which is beneficial in this context because it ensures consistency and relevance between the training and evaluation data. This alignment validates our evaluation approach and confirms that the synthetic data is appropriate and effective for training purposes.

W6.3 Both 1 and 2 are trying to identify the cause behind the performance degradation when both sources of data are combined in model fine-tuning. But regardless of the findings, table 3 and 8 do point to the need of a finer-categorization of the dataset. It may be helpful to separate the dataset into two parts and caution the users to combine them naively together.

Thank you for your valuable suggestion. We agree that the performance degradation observed when combining both data sources highlights the need for finer categorization of the dataset. We emphasized this point in Section 7 Limitations of the revised manuscript and explicitly cautioned users against naively combining the two fine-grained datasets during model fine-tuning. Your feedback is greatly appreciated and will help improve the clarity and utility of our work.

W6.4 Another evidence supporting the separation of interview data and synthetic data is the contrast between the context; The interview data was created based on a very focused group of caregivers offering care to a specific group of patients, this is in stark contrast to the unbounded target profile used in generating the synthetic data.

Thank you for your suggestion. This is a good reason that we should separate the two datasets apart. We included this observation in Section 7 Limitations in the revised manuscript to caution users about these contextual differences and remind them of the importance of handling the two datasets separately to avoid unintended biases or performance degradation.

W5. The findings in Table 3 and Table 8 also seem to point to questions left unanswered about the dataset. As can be seen in both of the tables:

W5.1 It is not true that more data generally leads to better performance

W5.2 It seems consistent that combining interview data with synthetic data actually HURTS model performance across all seven factors and model families

W5.3 It seems human evaluation consistently prefers models fine-tuned on synth data only

Thank you for your careful observation and comments. Indeed, more data may not be able to lead to better performance. As for the LLMs evaluation (GPT-4 and Gemini Pro), the synthetic data is generated by GPT-3.5 Turbo and there is an intrinsic bias where the GPT4 will prefer more the LLMs fine-tuned on the synthetic data, which indicates that the fine-tuning is effective. Therefore, all the models fine-tuned on synthetic data will get a higher score evaluated by GPT4. On the contrary, Gemini Pro prefers more the LLMs fine-tuned on the interview data, which may be due to its internal bias against ChatGPT (these two evaluators are competitors commercially). As a consequence, the LLMs fine-tuned on the combined data will not be preferred by either party.

As for the human evaluation, the models fine-tuned on synthetic data get the first place 5 times out of 7 cases. The models fine-tuned on combined data get the first place for 2 times out of 7 cases while they get the second place for 3 times out of 7 cases. This indeed reflects that humans prefer more on the models fine-tuned on synthetic data. However, in several cases, the interview data also plays an important role in making the models fine-tuned on the combined data have better performance.

All in all, since we only have 7 cases and the average scores or ranks are not significantly different among three kinds of models (i.e., fine-tuned on synthetic data only, interview data only, and combined data), the main conclusion we can draw is that with our MentalChat16K data, the fine-tuned models can be significantly improved in terms of both LLMs and humans evaluations.

W6. W5.3 alone could put into question the value of real data based on interviews if evaluation shows synthetic data seems to help model more. But the reviewer believes this may be attributed to misalignment between the interview data and the evaluation set. A couple of suggestions on analysis that could be included to pinpoint potential causes:

W6.1 Since the evaluation set contains data (InFamousCoder) originally designed for depression classification. Use the models (base and fine-tuned) to run classification on the question; the model performance could then point out the generalizability of the dataset (if fine-tuning benefits the model) or domain shift (if fine-tuning hurts the model performance) between MentalChat16k and the evaluation set.

Thank you for your suggestion. To address your question about generalizability and potential domain shift, we conducted a systematic evaluation using 100 randomly selected texts from the depression classification dataset. Each model was prompted with a structured classification task that required binary responses ("Yes"/"No") to indicate the presence of depression indicators. The results are shown in the table in the next comment (see Response to A2qn 5/6). Note that the two fine-tuned LLaMA2 did not follow the instruction to answer “Yes” or “No”, resulting in 0 accuracy.

Notably, most fine-tuned models performed comparably to or better than their corresponding base models in terms of accuracy, suggesting that the fine-tuning does not compromise the generalizability of the model in most cases. However, in instances where fine-tuned models showed degraded performance, this may be attributed to the task-specific nature of fine-tuning on mental health counseling rather than depression classification. While both tasks involve instruction-following, they differ in linguistic style and cues between MentalChat16k and the evaluation dataset. This mismatch could lead the model to misinterpret subtle indicators of depression, resulting in lower accuracy. Overall, fine-tuning on MentalChat16k preserves generalizability across tasks while enhancing the model's ability to follow structured instructions.

W2. Basic statistics about the dataset are lacking in the paper, I would recommend the following statistics be added and organized in a table:

Thank you for your suggestion. We have added an additional table to summarize the suggested dataset statistics in General Response 2/3. We also have added this information to Table 1 in Section 3.1 of our revised manuscript marked in blue.

W3 (Quality). It is unclear why the authors would break a complete interview, which the reviewer assumes to be a coherent series of questions and answers between healthcare coaches and caregivers, into individual QA pairs. It would be constructive to include discussion or reasoning on this particular choice for data collection. Specifically, Please explain how the loss of contextual information, when breaking conversation into QA pairs, would affect the data quality. Is the paraphrasing conducted to address partly this concern? So that each QA pair becomes somewhat independent of each other?

Thank you for your thoughtful question. The decision to break interviews into individual QA pairs was motivated by several considerations:

• Privacy Protection: Segmenting the interviews enables the removal of sensitive private information during paraphrasing.
• Improved Clarity: Fragmentary conversations were summarized into self-contained QA pairs to ensure uniformity in format and length.
• Practical Constraints: The large language model’s context length limitations make it infeasible to process or paraphrase entire documents (often exceeding 15 pages) into single QA pairs.

We acknowledge that breaking interviews into QA pairs may result in some loss of broader contextual information. To mitigate this, paraphrasing was designed to preserve the essence of each interaction while making the QA pairs as self-contained as possible. Additionally, the original order of the QA pairs is maintained, allowing downstream processes to reconstruct context if needed.

This approach balances the practical requirements of dataset preparation with the need to retain meaningful context. We have expanded the discussion of this point in Section 7 Limitations in the revised manuscript.

W4 (Quality). The introduction of the seven-factor evaluation metrics on empathy is interesting and novel to the reviewer, but because this is the only evaluation done by the authors, it also tends to conflate the focus of the paper into a paper of both a new dataset and a new, untested metrics. A few related points that could strengthen the paper, but lacking in the current version, are:

W4.1 Why are the scoring scale different between LLM-based and human evaluation?

In our study design, the LLM judges rate each response independently on a scale from 1 (poor) to 10 (excellent), while the human evaluators are asked to rank responses from different models, assigning ranks from 1 (best) to 7 (worst) for a given question. Using both methods complements the evaluation process. LLM scoring provides objective and granular insight into response quality, while human ranking captures intuitive and comparative judgments, minimizing bias and simplifying evaluation tasks. Together, they offer a holistic view of model performance.

We extend our sincere gratitude for dedicating your valuable time to reviewing our paper and for acknowledging the significance and originality of our contribution. We appreciate your suggestions on improving clarity and addressing potential issues related to combining datasets and evaluation methodologies. We have carefully considered your points and provide detailed responses below.

W1 (Clarity). A few details on the data collection procedure could be clarified. For example:

W1.1 Is the demographic information the only type of meta data collected?

Thank you for your question. Yes, in the original version of our paper, the demographic information (as shown in Table 4 in revised Appendix A.3.1) is the only type of meta-data collected for the interview data. We have provided an additional table that shows other metadata such as the size of the samples, the average QA pairs in each visit, etc. as you suggested in the following comments. For the details, please refer to Section A2 of General Response 2/3. This demographic information was gathered by the research group during the intervention study. For the synthetic data, we also have the topic distribution that was used to generate the data. In the revised manuscript, the detailed data table is presented in Table 1 in Section 3.1 and the topic distribution is detailed in Appendix A.3.2.

W1.2 What about the 93 caregivers without demographic information? Will the question-answer pairs from them be missing the meta data?

Thank you for your question. Yes, the data from the caregivers who did not provide demographic information will be missing that meta-data. The caregivers were given the option to not disclose their demographic details during the survey, and as a result, their responses will lack this specific information. However, the question-answer pairs themselves are still included in the dataset.

W1.3 How good (in terms of diversity, topic coverage, etc.) are the synthetic question-answer pairs?

Thank you for your question. The synthetic data covers 33 mental health topics, including Relationships, Anxiety, Depression, Intimacy, Family Conflict, etc., which covers all the main topics in the mental health domain [1] and their distributions are consistent with the real distributions reflected in the real world. Guaranteeing these two points makes the synthetic data not only comprehensive but also reflects the real situation, which indicates its superiority. The topic distribution has been included in Figure 3 in Appendix A.3.2 of our revised manuscript.

[1] American Psychology Association, www.apa.org.

W1.4 It would be great if the authors could provide concrete examples of the actual data, both from interviews and from the synthetic data.

Thank you for your suggestion. We agree that including some examples from the dataset could enhance the completeness of the paper and provide readers with a more intuitive understanding of the content. To address this, we incorporated a few representative examples of the synthetic data and the interview data in our revision (See Appendix A.11). We also attached the same examples in General Response 3/3. For details, please refer to that section.

W1.5 A brief explanation (or examples) that lays out the data structure (.json format? plain txt? what meta data is available? any data splits?) would be helpful, either in the main body or in the Appendix.

Thank you for your suggestion. The data is stored in .json format. For a summary of dataset statistics, please refer to Section A2 of General Response 2/3, where we provide detailed information about the dataset structure, including session counts, average QA pairs per session, word counts, and topic distributions. After the paper is accepted, we will provide a link to the dataset, including comprehensive documentation covering its structure, usage, and limitations. We appreciate your feedback and will ensure this information is clearly conveyed.

W6.5 If the goal of the dataset is to help improve empathetic responses from LLMs in the setting of mental healthcare counseling, then the reviewer suggests expanding the dataset by covering a wide range of responses different in level of empathy; that is, intentionally including non-empathetic responses, to better cover the range of the seven-factor metrics. Given the novelty of the metrics, such a dataset could also be ideal to further evaluate the robustness of the metrics and LLM's capability of automated evaluation. This could be an extension of future work.

Thank you for your insightful suggestion. Expanding the dataset to intentionally include responses with varying levels of empathy, including non-empathetic ones, is a great idea for enhancing the robustness of our metrics and further evaluating LLMs’ performance. While empathy is indeed a critical focus of our work, we want to highlight that it is just one of the seven metrics we employ to assess the functionality of a language model comprehensively. The remaining metrics—Active Listening, Safety & Trustworthiness, Open-mindedness & Non-judgment, Clarity & Encouragement, Boundaries & Ethical, and Holistic Approach—address broader aspects to ensure a balanced evaluation of a language model’s performance in mental health counseling. We will consider exploring this direction in future work to further strengthen the robustness and applicability of our dataset and metrics. Thank you again for your valuable input.

Q1. Do you intend to publish the dataset? It is unclear from the writing whether this is intended for public use. I apologize if I missed a statement or url on this.

Thank you for your question. Yes, we intend to publish both the synthetic and interview datasets, making them accessible to the broader research community. This will enable further advancements in mental health-related conversational AI research. We apologize if this was unclear in the manuscript and appreciate the opportunity to clarify.

Q2. Clarifying question: what is the hypothesis testing used in calculating P-value in Table 3 and Table 8? I assume it's two sample t-test, but it would be nice to clarify it in writing.

Thank you for your question. Yes, we used a two-sample t-test to evaluate the null hypothesis that the scores for the fine-tuned models and the base models have identical average (expected) values across the specified metrics. This statistical test was chosen to assess whether the observed differences in performance are statistically significant. We clarified this detail in Section 3.3 in the revision.

Q3. Are the 200 questions crawled from Reddit and Mental Health Forum to be included as part of the dataset, or are they mostly for evaluation purpose for this paper? The depression data from Reddit, judging by the dataset from InFamousCoder, are typically not in the form of questions (more like posters explaining the hardship they've been through), were there any filtering or cleaning involved in selecting questions from the data?

Thank you for the question. The 200 questions were collected specifically for evaluation purposes in this study and are not part of the MentalChat16K dataset. Regarding the depression data from Reddit, we acknowledge that such data often contains detailed narratives rather than explicit questions. To curate meaningful evaluation data, we applied a rigorous filtering and cleaning process to extract clear, self-contained questions from these narratives. This involved identifying and rephrasing key aspects of the posts that aligned with our evaluation metrics, ensuring relevance and clarity for use in model assessment.

M1. In Table 3 and 8, would it be better to rearrange the models in each family by the number of examples used in training? That is, order them by Base Model, Base Model + Interview data (7K), Base Model + Synth Data (10K), Base Model + Both (16K). It could help visualize the effect of dataset size on the finetuning (which is not significant, as would be obvious in the proposed order).

Thank you for your thoughtful suggestion regarding the organization of Tables 3 and 8 (now Tables 4 and 9 in the revised manuscript). We have updated these tables to arrange the models within each family in the proposed order: Base Model, Base Model + Interview Data (6K), Base Model + Synthetic Data (10K), and Base Model + Both (16K). This revised ordering better highlights the relationship between dataset size and the impact on fine-tuning performance, aligning with the point you raised about the limited effect observed. We appreciate your valuable feedback, which has helped us improve the clarity and interpretability of our results.

M2. Typos: page 2, "toassess"

Thank you for point out the typo. We fixed this typo in the revision.

Originally I thought the synthetic data or interview data would help models fine-tuned perform better on the original task (depression classification) of the evaluation data; but it seems the results are inconclusive: interview data seems to benefit model in depression classification in some cases, but there are also a lot of cases where fine-tuning doesn't help at all. I suspect this may be due to one or more reasons: (i) the questions in the evaluation set do exhibit a strong domain shift from the questions in the dataset, thereby brining into question the value of using InFamousCoder data for evaluation; (ii) the two tasks (generating responses vs depression classification) are different enough that the dataset is simply not generalizable to the other task. This unfortunately does seem to limit the use of the dataset.

We appreciate your thoughtful observation about the model performance on the depression classification task. Upon careful analysis, we identified that the initial performance issues were primarily related to the prompt design rather than the dataset's generalizability. Specifically:

1. Initial Prompt Analysis: Our original prompt was highly restrictive and formal but lacked vivid examples, which may reduces its effectiveness. Several LLMs cannot follow the instruction to succesffully answer any of the 100 questions:

Task: Depression Text Classification
Instructions:
1. Read the input text carefully
2. Classify whether the text indicates signs of depression
3. YOUR RESPONSE MUST BE EXACTLY ONE WORD:
   - Respond with "Yes" if depression indicators are present
   - Respond with "No" if depression indicators are not present
4. ANY OTHER RESPONSE FORMAT IS INCORRECT
5. DO NOT provide explanations, analysis, or additional commentary
6. DO NOT use punctuation marks

Format Check:
- Correct responses: "Yes" or "No"
- Incorrect responses: "Yes.", "No.", "Maybe", "The text shows depression", etc.

Input: [Text to analyze]
Required Output Format: [Yes/No]
### Response:

2. Improved Prompt Design: To handle this issue, following the successful approach demonstrated in "Towards Interpretable Mental Health Analysis with Large Language Models" (Zhang et al., EMNLP 2023), we adopted a few-shot prompt design that includes carefully selected examples. This design choice is supported by their findings that example-based prompting improves model performance in mental health analysis tasks. Our new prompt better aligns with real-world mental health conversations:

You will be presented with a text from Twitter. Classify whether the text indicate or include signs of depression.
- Respond with "Yes" if depression indicators are present
- Respond with "No" if depression indicators are NOT present

Here are some examples:
Question: Does the text indicate or include signs of depression?
Text: i want to be dead ive been suicidal for year im such a fucking retard filled with regret and anger im done with life i want all of this to stop why i cant do one thing right
Yes or No: Yes

Question: Does the text indicate or include signs of depression?
Text: i ve been with my grandma for day even she s about house away i ll still miss her
Yes or No: No

Question: Does the text indicate or include signs of depression?
Text: [text]
### Yes or No:

3. Improved Results: With this empirically-validated prompt design, our results demonstrate varied improvements across different fine-tuning configurations:

One more question about these seven metrics: are these a set of standard metrics used in mental health evaluation? If so, can you add reference to it (sorry if I missed it in related work)? If this is a set of metrics designed just for the dataset, then as I pointed out in my original comment: this would be a major weakness of the paper, that it conflates a new dataset and a new, untested metrics together in one paper, which may be very challenging to thoroughly analyze with convincing results.

Thank you for raising this important question regarding the evaluation metrics used in our study. We appreciate the opportunity to clarify that our seven metrics are indeed grounded in established therapeutic principles and practices, as well as recent research in AI-based mental health support evaluation. Our evaluation framework represents a synthesis of existing assessment approaches, including those from ChatCounselor's framework (Liu et al., 2023), and incorporates insights from recent works such as PsyEval (Li et al., 2023) and Towards Interpretable Mental Health Analysis with Large Language Models (Yang et al., 2023).

Each metric was carefully selected based on its validation in previous research and its relevance to effective therapeutic interactions. For example, our emphasis on empathy assessment aligns with the principles outlined in EPITOME (Sharma et al., 2020), while our focus on safety and trustworthiness draws from frameworks like Dialogue Safety Assessment (Qiu et al., 2023). By adapting these metrics for AI evaluation, we follow precedents set by studies such as PsyQA (Sun et al., 2021) and other recent contributions to the field.

Here is a detailed mapping of our metrics to their origins and supporting literature:

Our metrics framework thus represents a careful synthesis of established evaluation approaches in both traditional therapeutic practice and AI-based mental health support. Rather than being entirely new metrics, they are carefully adapted versions of validated measures, enhanced to address the specific challenges of AI evaluation in therapeutic contexts.

The effectiveness of these metrics is further validated by recent work such as PsyEval [9], which demonstrated the reliability of using GPT-4 for scoring model outputs on similar dimensions. This aligns with our approach while providing additional validation of the evaluation methodology.

We acknowledge that we could have made these connections to existing literature more explicit in our original submission and appreciate the opportunity to clarify these relationships. We have updated the paper to better highlight how our metrics build upon and integrate these established frameworks in Section 3.3 and Table 2 of the updated PDF.

References:

[1] Miller & Moyers (2006): Eight Stages in Learning Motivational Interviewing

[2] Sun et al. (2021): PsyQA: A Chinese Dataset for Generating Long Counseling Text

[3] Yang et al. (2023): Towards Interpretable Mental Health Analysis with Large Language Models

[4] Sharma et al. (2020): EPITOME: An Empathy-based Platform for Therapeutic Online Mental Health Education

[5] Rogers (1957): The Necessary and Sufficient Conditions of Therapeutic Personality Change

[6] Liu et al. (2023): ChatCounselor: A Large Language Model for Mental Health Counseling

[7] Qiu et al. (2023): Dialogue Safety Assessment with LLMs

[8] Lambert & Barley (2001): Research summary on the therapeutic relationship and psychotherapy outcome

[9] Li et al. (2023): PsyEval: A Suite of Mental Health Related Tasks for Evaluating Large Language Models

[10] Kabat-Zinn (2013): Full Catastrophe Living: Using the Wisdom of Your Body and Mind to Face Stress, Pain, and Illness

[11] American Psychological Association Ethics Code (2017) (edited)

Q2: Would the paper benefit from including side-by-side examples of model responses versus human counselor and baseline model responses? Could such examples better illustrate the model’s performance on key metrics like empathy, clarity, and safety?

Thank you for this thoughtful suggestion about including comparative examples. While we agree that examples could effectively illustrate the differences in response quality and help readers understand how we evaluated key metrics, we did not include any examples in the manuscript due to length constraints. As evident from the examples below, each response is quite lengthy, and including multiple sets would significantly increase the paper's length. We included the following examples in Appendix A.11.

This example includes a question-answer pair from the Interview Data and responses to the input question generated by Mistral-Instruct-V0.2 and its fine-tuned version fine-tuned on MentalChat16K. The human counselor and fine-tuned model both acknowledge the emotional experience ("I can see that you're experiencing a great deal of distress" vs "I'm truly sorry to hear that you've been experiencing such intense emotions"), while validating feelings and the experience. All responses appropriately address mental health concerns while staying within ethical boundaries. They encourage professional help when needed and avoid potentially harmful advice.

Input: I've been feeling so overwhelmed lately. I can't seem to focus on anything, and I've been having panic attacks almost daily. I've tried to reach out to friends and family, but they don't seem to understand what I'm going through. I feel so alone and lost.

Paraphrased Ground Truth Response from Human: I can see that you're experiencing a great deal of distress and feeling quite isolated. It's important to acknowledge the intensity of your emotions and the impact they're having on your daily life. It's not uncommon for individuals dealing with anxiety and depression to feel misunderstood by those around them. In fact, research shows that people with mental health conditions are often stigmatized and misunderstood by society. However, it's essential to remember that you're not alone in this experience. Many people face similar challenges, and there are resources and support systems available to help you.

As we work together in this therapeutic process, I'd like to propose a few strategies to help you manage your anxiety and improve your overall well-being. First, we can explore cognitive-behavioral techniques to help you challenge and reframe negative thoughts that contribute to your anxiety and panic attacks. Second, we can work on developing coping skills, such as deep breathing exercises, progressive muscle relaxation, and mindfulness meditation, to help you manage your symptoms in the moment. Lastly, we can explore the possibility of connecting you with support groups and other resources in your community to help you build a network of understanding and compassionate individuals who can provide you with the emotional support you need.

It's important to remember that healing is a process, and it takes time and effort to make meaningful progress. But with a compassionate and supportive therapeutic environment, I believe you can overcome these challenges and find the peace and understanding you're seeking.

W3. Outdated Evaluation Techniques: The paper employs simple prompt-based evaluation methods but does not leverage advanced frameworks in NLG evaluation. State-of-the-art methods such as G-Eval [1] and Prometheus [2] provide more sophisticated and reliable assessments of LLM outputs, which could yield a more accurate and nuanced evaluation of the models’ effectiveness. [1] G-Eval: NLG Evaluation using GPT-4 with Better Human Alignment https://arxiv.org/abs/2303.16634 [2] Prometheus: Inducing Fine-grained Evaluation Capability in Language Models https://arxiv.org/abs/2310.08491

Thank you for your insightful feedback and for bringing attention to G-Eval and Prometheus. Both frameworks represent significant advancements in NLG evaluation by providing more fine-grained, reliable, and nuanced assessments of LLM outputs. In our current study, we opted for prompt-based evaluation methods such as FastChat [1] due to their simplicity, interpretability, and alignment with the specific goals of validating the proposed dataset. While these methods effectively serve our research objectives, we recognize that the state-of-the-art frameworks like G-Eval and Prometheus could offer additional dimensions of analysis and enhance the depth of our evaluations.

We have included references to these advanced methods in our revised paper and recommend them as potential alternatives for users seeking more comprehensive evaluation tools in Section 7 Limitations. Additionally, we will consider adopting such methodologies in future research to further strengthen the reliability and scope of our evaluation approach.

[1] Zheng, Lianmin, et al. "Judging llm-as-a-judge with mt-bench and chatbot arena." Advances in Neural Information Processing Systems 36 (2023): 46595-46623.

Q1. How were human evaluators instructed to rank model responses, and were there specific criteria or guidelines provided to ensure consistency and objectivity? Additionally, was any inter-rater reliability measure (e.g., Cohen’s Kappa) applied to validate the consistency of human rankings across evaluators?

Thank you for your question. For the first question, we instructed human evaluators to rank model responses based on the effectiveness of the answer in addressing the given input. Effectiveness was defined as the response's relevance, empathy, clarity, and ability to appropriately address the concerns.

For the second question, we acknowledge that we did not include inter-rater reliability measures in our evaluation due to limited resources. However, to mitigate potential biases, we made an effort to include 4 human evaluators, ensuring that the evaluation results were not overly influenced by a single evaluator's perspective. To address consistency in rankings, we calculated the inter-rater reliability using Cohen’s kappa score. The average Cohen’s kappa score across all pairs of evaluators was 0.441, indicating “moderate agreement” based on the Landis and Koch (1977) scale. This suggests a reasonable level of alignment between evaluators while also highlighting areas for potential improvement. While the current kappa score reflects moderate agreement, we recognize the importance of higher inter-rater reliability to ensure the robustness of our findings. Moving forward, we plan to refine the evaluation process by incorporating additional training, clearer evaluation rubrics, and more rigorous inter-rater reliability assessments to further enhance consistency where resources and time permits. Please refer to Section A1 of General Response 2/3 for more details. The details are also discussed in Section 3.3 from Lines 312 to 323 in the revised manuscript.

We sincerely appreciate your feedback, as it will guide our efforts to improve methodological rigor in future evaluations.

We sincerely appreciate your insightful review and your recognition of our efforts in developing and fine-tuning LLMs for mental health counseling. We have thoroughly addressed your concerns and hope that the main concerns raised in the review can be clarified by the following responses.

W1. Reliability of LLM-Based Evaluation: The study primarily uses GPT-4 Turbo and Gemini Pro as automated evaluators to assess the performance of fine-tuned models. This approach presumes that these LLMs can objectively and accurately evaluate counseling responses across seven custom-defined mental health metrics. However, due to inherent biases and limitations, LLMs may not be entirely reliable as evaluators, especially in specialized fields such as mental health counseling.

We appreciate this concern about evaluation reliability. Our approach to evaluation was carefully designed with multiple layers of validation:

• We deliberately employed two distinct and state-of-the-art LLMs (GPT-4 Turbo and Gemini Pro) for evaluation, as divergent architectures help mitigate model-specific biases.
• This methodology of using powerful third-party LLMs for evaluation is well-established in the field. There are many papers that established the precedent and validity of using third-party LLMs as evaluators including several notable papers such as Vicuna [1], PaLM2 [2], Self-Instruct [3], etc.
• To further strengthen our evaluation framework, we supplemented the LLM evaluations with human expert assessments. The agreement between all the human evaluators (Cohen’s Kappa score 0.441) provides additional validation of our evaluation methodology. While we acknowledge that no evaluation method is perfect, our multi-model, human-validated approach provides a robust framework for assessing counseling responses, aligning with current best practices in the field.

[1] Chiang, Wei-Lin, et al. "Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality." See https://vicuna.lmsys.org (accessed 14 April 2023) 2.3 (2023): 6.

[2] Anil, Rohan, et al. "Palm 2 technical report." arXiv preprint arXiv:2305.10403 (2023).

[3] Wang, Yizhong, et al. "Self-instruct: Aligning language models with self-generated instructions." arXiv preprint arXiv:2212.10560 (2022).

W2. Inconsistency in Evaluation Scores: As shown in Table 3, there is a notable inconsistency between evaluations from GPT-4, Gemini Pro, and human experts. Almost all scores from Gemini are higher than those from GPT-4, and Gemini’s scores diverge significantly from Human Rankings. This discrepancy raises concerns about the validity and robustness of using LLMs as evaluators, as they might not effectively capture the nuances involved in mental health counseling.

Thank you for this observation. While there are differences in absolute scores between evaluators, it's crucial to note several key points:

• Most importantly, all three evaluation methods (GPT-4, Gemini Pro, and human experts) consistently demonstrate that fine-tuned LLMs outperform their base versions. This unanimous agreement across different evaluators strongly validates the effectiveness of our MentalChat16K dataset for improving LLMs' mental health counseling capabilities, which is the primary objective of our dataset paper.
• The score differences between GPT-4 and Gemini Pro can be attributed to the phenomenon that GPT-4 tends to show bias toward responses generated by GPT-3.5, likely due to their architectural similarities and training data overlap as discussed in Section 4.5. Gemini Pro, being developed independently with different architecture and training data, provides a more independent evaluation perspective.
• The human evaluation results align particularly well with Gemini Pro's assessment, suggesting that Gemini's scores might better reflect human judgment in this domain. This alignment adds credibility to our evaluation framework, as human expert judgment is considered the gold standard.

To avoid confusion in interpreting the results, we have added directional indicators (↑/↓) to Tables 4 and 9. The opposite directions stem from different evaluation strategies: human evaluators use rankings (where lower values indicate better performance), while GPT-4 and Gemini Pro use numerical scores (where higher values indicate better performance). Despite these different scoring conventions, all evaluators consistently show that fine-tuning improves model performance.

The consistent improvement trend across all evaluators, despite their methodological differences, provides strong evidence for the effectiveness of MentalChat16K in enhancing LLMs' mental health counseling capabilities.

Thank you for your thoughtful and detailed responses, particularly regarding the consistency of human evaluations and the ethical considerations. I appreciate the clarifications you've provided, which have helped address my concerns.

However, my primary remaining concern lies with the evaluation method used in the study. Specifically, I am worried about the reliability of prompt-based evaluation methods. Given the rapid evolution of commercial models and the potential for specific prompts to lose effectiveness as the models are iterated upon, I am concerned about the reproducibility of these evaluations over time. In contrast, methods like Prometheus, which fine-tune specific open-source models, mitigate this issue, as they are less dependent on potentially unstable prompts.

If you can address this concern and clarify how you ensure the stability and reproducibility of your evaluation approach, I would be willing to reconsider my score.

Thank you for your detailed responses, which addressed part of my concerns. I am happy to raise my rating to 5. However, I share Reviewer A2qn’s concerns about the seven metrics used in the evaluation. Specifically, it is important to clarify whether these metrics are standard in mental health evaluation or if they were specifically designed for this dataset. As Reviewer A2qn noted, if these metrics are newly proposed, it raises questions about their generalizability and validation.

Paraphrased transcript (one QA pair)

Caregiver: Expressing the importance of physical touch and its impact on my emotional wellbeing, I've found that receiving daily hugs significantly improves my mood and overall mental health. However, when my 'hug tank' gets low, I can become ill. I'm grateful for the mutual support group I've joined, where I receive hugs from my loved one daily.

Interviewer: Empathizing with your experience, I understand the power of human connection and touch in maintaining mental health. It's essential to prioritize and seek out opportunities for physical affection, as it can have a profound impact on our emotional wellbeing. Additionally, when you're feeling low on hugs, consider reaching out to friends, family, or even strangers for a hug to help replenish your 'hug tank'. Remember, it's important to take care of yourself and prioritize your emotional needs.

To provide additional transparency, we have expanded the discussion in Section 7 Limitations of the paper to explicitly acknowledge these rare instances of deviation and the inherent challenges associated with paraphrasing. We have also included further examples in Appendix A.11 to highlight both successful paraphrasing and the specific contexts where minor inconsistencies may arise. This addition will offer readers a clearer understanding of the paraphrasing process and its potential limitations.

W4. Inter annotation agreement for human evaluation would enhance the human evaluation.

Thank you for your valuable suggestion. To mitigate potential biases, we included four human evaluators to ensure the evaluation results were not overly influenced by a single evaluator’s perspective. We recognize the importance of inter-annotator agreement as a metric for assessing the consistency and reliability of human evaluations.

To address this concern in the current study, we calculated the inter-annotator agreement using Cohen’s kappa score. The average Cohen’s kappa score across all pairs of evaluators was 0.441, which falls in the “moderate agreement” range based on the scale by Landis and Koch (1977) [1]. This demonstrates a reasonable level of consistency among the evaluators. For details, please refer to Section A1 of the General Response 2/3. The details are also discussed in Section 3.3 from Lines 312 to 323 in the revised manuscript.

We appreciate your feedback and will strive to enhance our use of inter-annotator agreement strategies in future studies, where resources and time permit.

[1] Landis, J. R., & Koch, G. G. (1977). The measurement of observer agreement for categorical data. Biometrics, 33(1), 159-174.

We are honored to have you dedicate your valuable time to reviewing our paper and for expressing your appreciation and support for our work. Your insights have significantly enhanced the clarity and depth of our manuscript. In the following, we have carefully considered your suggestion and provide a comprehensive response to your review comments.

W1. The paper should have had a table summarising the dataset statistics for easy access.

Thank you for your suggestion. We have included a table (Table 1) in Section 3.1 that summarizes the statistics of our data marked in blue. We also put the table in General Response. For more details, please refer to Section A2 of General Response 2/3.

W2. Examples of generated counseling transcripts in the paper would be useful to evaluate them.

Thank you for your suggestion. We agree that including some examples from the dataset could enhance the completeness of the paper and provide readers with a more intuitive understanding of the content. To address this, we incorporated a few representative examples of the synthetic data and the interview data in our revision (See Appendix A.11). We also attached the same examples in General Response 3/3. For details, please refer to that section.

W3. It would be interesting to observe if the paraphrasing step, as illustrated in Table 1, results in any hallucinations to complete the given "requirements" in the prompt.

Thank you for your insightful observation. We agree that examining whether the paraphrasing step results in hallucinations is an important aspect of evaluating the robustness and reliability of our approach. We did conduct preliminary analyses to assess the fidelity of paraphrased outputs against the original conversation transcript, with assistance from the research group that conducted the intervention study. Specifically, we reviewed a subset of paraphrased examples for consistency with the given requirements and found that the model generally adheres to the intended meaning without introducing significant hallucinations. However, as with most generative tasks, we acknowledge that some minor deviations were observed in rare cases, typically in complex or ambiguous prompts. These deviations were infrequent and typically involved subtle rephrasing rather than significant hallucinations. To illustrate this point, the Table below compares an example from an original transcript with its paraphrased counterpart. This example demonstrates how the paraphrased version maintains the core message and context of the original dialogue while adapting it to a concise QA format.

Original transcript (One page) (BHC stands for Behavioral Health Coach)

Caregiver: Definitely. We’re, a mutual support group.

BHC: [laughs] Good for you. Uh, it makes all the difference.

Caregiver: It really does. It really does and that way you get your, uh, daily quota of hugs, too –

BHC: Yeah.

Caregiver: - if you have somebody who – who is coming and going. I get a hug every time he, uh, comes home from work and when he goes off.

BHC: Oh, it’s so important – that touch. Don’t you think?

Caregiver: If – if your hug tank gets too low, that’s when you can also make yourself ill.

BHC: Oh, you’re so wise. Sandra, you- I’m so glad you’re telling us all this. It’s really helpful to us. I’m making notes.

Caregiver: Only who knows and – you know, they’re-, well, people have become-, when I was growing up, people weren’t as hug-y as they are now.

BHC: Uh-huh [affirmative]?

Caregiver: Uh, and I’m so glad that they are. Even those who, um, my-, my friend who just died on Tuesday – was, uh, grew up in the East and it was a little more, um, uh, I don’t know – proper? Uh, her family was not demonstrative. And so it was a long time ‘til we could give her a hug and it was always a very gentle, just kind of on the shoulder, so just a little squeeze, but she knew that we loved her.

BHC: Mm-hmm.

Caregiver: Uh, and so – but then there are the bear-hug people. Oh, we could talk for a half-hour just on huggers.

BHC: [laughs] Oh, good. I’m so glad –

Caregiver: [laughs]

BHC: - so glad to hear that you know what you need and you take full advantage of the opportunities available to you [laughs].

Caregiver: That’s for sure. I – if necessary – if I have to, if it gets to that, then I’ll go up to strangers. [laughs]

BHC: That’s great, Sandra. I’m so pleased. I’m – OK. Um, all right. Well, next time we’ll-, we’ll choose a concern and we’ll kinda break it down and well do the D in ADAPT – which is defining. Um –

Caregiver: All right.

For the paraphrased transcript (one QA pair) and the remaining responses, please see Response to Reviewer VT66 2/2.

We extend our sincere gratitude for dedicating your valuable time to reviewing our paper and for expressing your appreciation and support for our work. We are delighted that you recognized the significance of MentalChat16K as a valuable benchmark dataset for advancing conversational AI in mental health counseling. In the following, we have provided a comprehensive response to each of your review comments.

W1. Who annotated MentalChat16K (referred to as human experts in lines 182 and 188)? Are these the same annotators who evaluated model performance (referred to as human experts in line 275)?

Thank you for pointing this out. These are two different groups of human experts. In lines 182 and 188, the human experts refer specifically to the behavioral health professionals who transcribed the interview recordings to text. The evaluators in line 275 were independently selected Postdocs, PhD, and Masters’ students who provided performance assessments.

W2. MentalChat16K contains a total of 6,338 question-answer pairs. How many conversations does the dataset comprise? Are the pairs presented as individual exchanges, or do they maintain conversational continuity?

Thank you for your question. The MentalChat16K dataset comprises a total of 378 unique conversation sessions, each containing multiple question-answer pairs that maintain conversational continuity. Each conversation represents a realistic dialogue flow between a patient and a counselor, rather than isolated exchanges. This structure captures the dynamics of real counseling sessions, and the detail of the paraphrasing process is illustrated in Section 3.1.1.

W3. What are the scores of the ground-truth responses in MentalChat16K when evaluated by the same metrics used to assess LLM performance? Since the evaluation method is reference-free, it would be useful to know how the dataset responses score according to the proposed metrics, and whether these scores exceed those of the fine-tuned LLMs.

Thank you for your thoughtful question and suggestion. This is indeed an important aspect to explore. We conducted an evaluation of the ground-truth responses in a randomly selected subset of 200 question-answer pairs from the MentalChat16K dataset (100 from Synthetic data and 100 from Interview data). Using the same metrics proposed for assessing LLM performance, we evaluated the ground-truth responses alongside the responses generated by 31 LLMs. The scores for the 7 metrics assigned to the ground-truth responses by Gemini-Pro are as follows:

It’s worth noting that the ground truth responses scored highest in Empathy & Validation and Open-mindedness & Non-judgment, which are key attributes for human mental health counseling.

When compared to the scores of LLM-generated responses, the average score of the ground-truth responses (7.95) ranked 12th among the 31 LLMs. Among the 11 LLMs that achieved higher average scores, 7 were fine-tuned on Interview data, 2 on Synthetic data, and 2 on both. Notably, the average score of the ground-truth responses exceeded all 10 base and baseline models.

W4. What test dataset was used in the experiments? Was an 80/20 split applied for training and testing, or was a different setup used? The reported p-values suggest multiple experimental runs, but Section 4.4 lacks details on the experimental setup.

Thank you for highlighting the need for clarity. In our experiments, we used a separate test dataset comprising 200 mental health-related questions collected from Reddit and mental health forums, which were not part of the training data. The MentalChat16K dataset was used exclusively for training purposes, and we did not apply an 80/20 split within it. The reported p-values result from five experimental runs to ensure robustness. We mentioned the details in Section 4.5 from lines 472 to 480 in the original manuscript. We have also added this detail to Section 3.3 from lines 300 to 303 in the updated manuscript to make it clearer.