File size: 7,192 Bytes
9203391
 
 
 
 
9f553e1
 
9203391
 
 
9f553e1
9203391
 
 
 
 
 
e69b9a8
 
9203391
 
 
 
9f553e1
 
 
 
 
 
9203391
 
 
341e9b1
9203391
 
 
 
 
 
 
 
 
 
9f553e1
9203391
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9f553e1
 
 
 
 
9203391
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
15b53fd
9203391
 
 
 
 
 
9f553e1
9203391
 
 
 
 
9f553e1
9203391
 
 
9f553e1
 
 
 
 
 
 
9203391
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9f553e1
 
9203391
 
 
 
 
 
 
9f553e1
9203391
 
 
 
 
 
 
 
 
 
 
 
e69b9a8
9203391
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
---
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",
}
```