--- license: apache-2.0 language: - zh library_name: transformers pipeline_tag: text-generation base_model: Qwen/Qwen3-8B base_model_relation: finetune tags: - counseltraj - counseltrajlm - counseling - conversational - qwen3 - merged --- # CounselTrajLM — Qwen3-8B 📄 **Paper:** [arXiv:2611.24168](https://arxiv.org/abs/2611.24168) CounselTrajLM generates the next Chinese counselor response from a counseling dialogue history. This version is fine-tuned on [CounselTrajD](https://github.com/UEC-InabaLab/CounselTraj#counseltrajd), using [Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) as its base. The repository contains the **complete merged BF16 model**, in nine safetensors shards totaling approximately **16.4 GB**, together with its tokenizer, chat template, inference script, and checksums. It loads directly with Transformers. Inference uses **non-thinking mode** (`enable_thinking=False`). This response-generation model is separate from the [CounselTraj trajectory planner](https://huggingface.co/UEC-InabaLab/CounselTraj). Project, dataset, and evaluation prompts are available in the [CounselTraj repository](https://github.com/UEC-InabaLab/CounselTraj). ## Training details | Item | Value | | --- | --- | | Base model | Qwen/Qwen3-8B | | Base revision | `b968826d9c46dd6066d109eabc6255188de91218` | | Training data | CounselTrajD; dialogue-disjoint 90/10 train/validation split | | Training / validation examples | 425,679 / 47,528 next-response examples | | Fine-tuning method | QLoRA; adapter merged into the unquantized BF16 base for release | | LoRA rank / alpha / dropout | 32 / 64 / 0.05 | | Learning rate | 0.0002 | | Schedule / warmup ratio | Cosine / 0.06 | | Epochs | 3; final-epoch export | | Final optimizer step | 79,815 | | Effective batch size | 16 | | Maximum training sequence length | 2,048 tokens | | Maximum history length | Last 1,800 characters | | Thinking mode | Disabled | | Release format | Complete merged BF16 safetensors | Validation was used for monitoring. The released model is the final epoch. See [training_summary.json](training_summary.json) for the recorded settings. ## Generate a reply Download the public model repository and install the inference dependencies: ```bash pip install huggingface_hub python -c 'from huggingface_hub import snapshot_download; snapshot_download("UEC-InabaLab/CounselTrajLM-Qwen3-8B", local_dir="CounselTrajLM-Qwen3-8B")' cd CounselTrajLM-Qwen3-8B pip install -r requirements.txt python generate_reply.py --model . --input-file examples/histories.jsonl ``` Or provide 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。' ``` Input JSONL rows contain an optional `id` and a nonempty `history` string. Use `client:` and `counselor:` speaker prefixes within the history. The script uses the original counselor system instruction and training prompt format, retaining the last 1,800 history characters. It outputs JSON with the generated `response`. The three bundled histories are newly written synthetic examples. After downloading the repository, it can also be used from Python: ```python from generate_reply import load_model, generate_reply tokenizer, model = load_model("UEC-InabaLab/CounselTrajLM-Qwen3-8B") response = generate_reply( tokenizer, model, "client:最近工作压力很大,晚上总是睡不好。", ) print(response) ``` The complete weights load in BF16, with additional memory required for inference. A separate base-model download or PEFT adapter loading is not required. The release was checked with Transformers 5.13.0, PyTorch 2.5.1, Accelerate 1.13.0, and safetensors 0.8.0. The GPU generation checks and integrity results are recorded in [validation.json](validation.json); file checksums are in [SHA256SUMS](SHA256SUMS). ## Training data and references CounselTrajD contains 10,964 synthetic Chinese counseling dialogues expanded from three sources: - SmileChat: Qiu et al. (2024), [SMILE](https://aclanthology.org/2024.findings-emnlp.34/). - PsyDTCorpus: Xie et al. (2025), [PsyDT](https://aclanthology.org/2025.acl-long.55/). - CPsyCounD: Zhang et al. (2024), [CPsyCoun](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. The model is released under Apache-2.0; see [LICENSE](LICENSE). ## Intended use and limitations The model is intended for research on Chinese counseling dialogue generation. Its training data are synthetic. Generated responses may be inaccurate, inappropriate, or unsafe, and the model is not a substitute for professional mental health care. ## 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", } ```