--- language: - zh base_model: Qwen/Qwen3.5-9B base_model_relation: finetune library_name: transformers license: apache-2.0 pipeline_tag: text-generation tags: - qwen3_5 - merged - psychological-counseling - counselor-response-generation --- # CounselTrajLM 📄 **Paper:** [arXiv:2611.24168](https://arxiv.org/abs/2611.24168) CounselTrajLM generates the next Chinese counselor response from a counseling dialogue history. It is fine-tuned on CounselTrajD and is the dialogue model used in the CounselTraj response-level and dialogue-level experiments. This release contains the **complete merged BF16 text language model**, its tokenizer, and a working inference script. It combines the previously released three-epoch LoRA adapter with the unquantized [Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) base at revision `c202236235762e1c871ad0ccb60c8ee5ba337b9a` and loads directly with Transformers. The model is released under Apache-2.0; see [LICENSE](LICENSE). Project, dataset, and evaluation prompts: [UEC-InabaLab/CounselTraj](https://github.com/UEC-InabaLab/CounselTraj). The separate [CounselTraj](https://huggingface.co/UEC-InabaLab/CounselTraj) model plans dialogue-act trajectories for data synthesis. ## Training details The following settings come from the saved training records for the released checkpoint. | Item | Value | | --- | --- | | Base model | Qwen/Qwen3.5-9B, text language-model component | | Release format | Complete merged BF16 text language model | | Fine-tuning | QLoRA, NF4 with double quantization and bfloat16 computation | | Adapter rank / alpha / dropout | 32 / 64 / 0.05 | | Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj | | Adapter parameters | 58,195,968 | | Dataset split | Dialogue-disjoint 90% training / 10% validation | | Training / validation examples | 425,679 / 47,528 next-response examples | | Training epochs | 3 | | Final optimizer step | 79,815 | | Learning rate / warmup ratio | 0.0002 / 0.06 | | Learning-rate schedule | Cosine | | Optimizer | paged_adamw_8bit | | Maximum training sequence length | 2,048 tokens | | Maximum history length | Last 1,800 characters | | Per-device training batch size / gradient accumulation | 2 / 1 | | Random seed | 20260704 | | Recorded final validation loss | 1.5481888056 | The merged model uses the final adapter saved after three epochs. Its SHA256 was verified against both the source final adapter and the previously published adapter. The trainer record contains no best-validation checkpoint selection. The [original adapter release](https://huggingface.co/UEC-InabaLab/CounselTrajLM/tree/341e9b198c985ee75c982a43376015d806b15522) remains available in the repository history. ## Training data and references [CounselTrajD](https://github.com/UEC-InabaLab/CounselTraj#counseltrajd) contains 10,964 synthetic Chinese counseling dialogues expanded from three sources: - SmileChat: Qiu et al. (2024), [SMILE: Single-turn to Multi-turn Inclusive Language Expansion via ChatGPT for Mental Health Support](https://aclanthology.org/2024.findings-emnlp.34/). - PsyDTCorpus: Xie et al. (2025), [PsyDT: Using LLMs to Construct the Digital Twin of Psychological Counselor with Personalized Counseling Style for Psychological Counseling](https://aclanthology.org/2025.acl-long.55/). - CPsyCounD: Zhang et al. (2024), [CPsyCoun: A Report-based Multi-turn Dialogue Reconstruction and Evaluation Framework for Chinese Psychological Counseling](https://aclanthology.org/2024.findings-acl.830/). See the project's [licensing scope](https://github.com/UEC-InabaLab/CounselTraj/blob/master/LICENSING.md) and [source notices](https://github.com/UEC-InabaLab/CounselTraj/blob/master/THIRD_PARTY_NOTICES.md) for the source-specific data terms. ## Generate a reply Download the public complete model and run the included example: ```bash pip install huggingface_hub python -c 'from huggingface_hub import snapshot_download; snapshot_download("UEC-InabaLab/CounselTrajLM", local_dir="CounselTrajLM")' cd CounselTrajLM pip install -r requirements.txt python generate_reply.py --model . --input-file examples/histories.jsonl ``` Or supply a history directly: ```bash python generate_reply.py --model . --history 'client:最近工作压力很大,晚上总是睡不好。' ``` The script uses the original counselor system prompt and user-history format, with `enable_thinking=False`. It loads the complete text language model in BF16, with additional memory required for the generation cache. A separate base-model download and the PEFT library are no longer required. The weights occupy approximately 17.9 GB. The exported model was reloaded on CUDA and checked on the three bundled synthetic histories. The merge comparison and file integrity results are recorded in [validation.json](validation.json). Input JSONL rows contain an optional `id` and a nonempty `history` string. The script outputs JSON with the generated `response`. The three bundled histories are newly written synthetic examples. ## Load in Python After downloading the repository and installing its requirements: ```python from generate_reply import load_model, generate_reply tokenizer, model = load_model("UEC-InabaLab/CounselTrajLM") response = generate_reply( tokenizer, model, "client:最近工作压力很大,晚上总是睡不好。", ) print(response) ``` ## Evaluation The response-level evaluation uses PsyDial-D101 from Qiu and Lan (2025), [PsyDial: A Large-scale Long-term Conversational Dataset for Mental Health Support](https://aclanthology.org/2025.acl-long.1049/). The dialogue-level evaluation uses client profiles from Pan et al. (2026), [PsychEval: A Multi-Session and Multi-Therapy Benchmark for High-Realism AI Psychological Counselor](https://arxiv.org/abs/2601.01802). The [evaluation documentation](https://github.com/UEC-InabaLab/CounselTraj/tree/master/evaluation) provides the prompts, dimensions, and research foundations. ## Files - `model-*.safetensors`, `model.safetensors.index.json`, and `config.json`: complete merged text language model. - `generation_config.json`: generation defaults. - `tokenizer.json`, `tokenizer_config.json`, and `chat_template.jinja`: saved tokenizer and chat template. - `generate_reply.py`: inference entry point. - `requirements.txt`: inference dependencies. - `model_details.json`: actual training settings and checkpoint provenance. - `examples/histories.jsonl`: synthetic example inputs. - `validation.json`: release integrity and generation checks. - `SHA256SUMS`: file checksums. - `LICENSE`: Apache-2.0 license. The model is intended for counseling dialogue research. Its generated responses have not been validated as clinical care or crisis support. ## Citation ```bibtex @inproceedings{qi-inaba-2027-counseltraj, title = "{CounselTraj}: Synthesizing Process-Aware Psychological Counseling Dialogues via Dialogue-Act Trajectory Planning", author = "Qi, Zhiyang and Inaba, Michimasa", booktitle = "Proceedings of the 20th Conference of the European Chapter of the Association for Computational Linguistics", year = "2027", url = "https://arxiv.org/abs/2611.24168", } ```