ZhiyangQi97 commited on
Commit
b6a1960
·
verified ·
1 Parent(s): a33e079

Release CounselTraj DA trajectory planner (Qwen3-8B LoRA)

Browse files
.gitattributes CHANGED
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ tokenizer.json filter=lfs diff=lfs merge=lfs -text
README.md ADDED
@@ -0,0 +1,124 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language:
3
+ - zh
4
+ base_model: Qwen/Qwen3-8B
5
+ base_model_relation: finetune
6
+ library_name: peft
7
+ pipeline_tag: text-generation
8
+ tags:
9
+ - qwen3
10
+ - lora
11
+ - dialogue-act-planning
12
+ - psychological-counseling
13
+ ---
14
+
15
+ # CounselTraj-DA
16
+
17
+ CounselTraj-DA is the dialogue-act trajectory planning model used in CounselTraj.
18
+ Given a brief counseling dialogue, the planner produces an ordered trajectory
19
+ of role-tagged acts such as `counselor:[OpenQuestion]` and
20
+ `client:[SelfDisclosure]` for subsequent dialogue synthesis.
21
+
22
+ This repository contains a **LoRA adapter for
23
+ [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B)**, together with its tokenizer,
24
+ the trajectory planning script, and aggregate dialogue-act priors. Loading the
25
+ adapter also requires the Qwen3-8B base model. Its base model is distributed under
26
+ Apache-2.0.
27
+
28
+ Project and dataset: [UEC-InabaLab/CounselTraj](https://github.com/UEC-InabaLab/CounselTraj).
29
+
30
+ ## Model details
31
+
32
+ | Item | Value |
33
+ | --- | --- |
34
+ | Base model | Qwen/Qwen3-8B |
35
+ | Adapter | LoRA, rank 8, alpha 16, dropout 0.05 |
36
+ | Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
37
+ | Adapter parameters | 21,823,488 |
38
+ | Selected checkpoint | checkpoint-145, epoch 5 |
39
+ | Selection criterion | Lowest recorded validation loss, 1.1678476334 |
40
+ | Learning rate / warmup ratio | 0.0002 / 0.03 |
41
+ | Planner | Content-conditioned v4 constrained decoding |
42
+ | Default trajectory length budget | 64–92 dialogue acts |
43
+
44
+ The released adapter is byte-identical to both the selected checkpoint and the
45
+ saved final model. Training uses dialogue-act annotations of counseling
46
+ dialogues; the bundled stage, transition, and global priors were aggregated from
47
+ 1,010 annotated dialogues.
48
+
49
+ ## Generate a trajectory
50
+
51
+ Authenticate with a Hugging Face account that has access to this private model:
52
+
53
+ ```bash
54
+ pip install huggingface_hub
55
+ hf auth login
56
+ python -c 'from huggingface_hub import snapshot_download; snapshot_download("UEC-InabaLab/CounselTraj-DA", local_dir="CounselTraj-DA")'
57
+ cd CounselTraj-DA
58
+ pip install -r requirements.txt
59
+ python plan_trajectory.py --input-file examples/input.jsonl --limit 1
60
+ ```
61
+
62
+ The default inference path uses CUDA and bitsandbytes NF4 quantization. On a
63
+ supported device with enough memory, `--no-4bit` loads the base model without
64
+ 4-bit quantization.
65
+
66
+ Input is UTF-8 JSONL with an `input` string containing the brief dialogue. Optional
67
+ fields include `id`, `topic`, `topic_zh`, `source_dataset`, and `source_id`. The
68
+ bundled example is a newly written synthetic illustration. The output JSONL
69
+ contains the planned role-tagged trajectory in `output`, along with its length
70
+ and tag counts.
71
+ The original planner can append closing acts beyond its target length budget;
72
+ count the acts in `output` when an exact realized length is needed.
73
+
74
+ The planner combines adapter likelihoods with stage, transition, and global
75
+ priors, content and phase scores, a repetition penalty, and a diversity bonus.
76
+ Use `plan_trajectory.py` to run this decoding procedure. Loading the adapter alone
77
+ provides the language-model component of the planner.
78
+
79
+ ## Load the adapter in Python
80
+
81
+ ```python
82
+ import torch
83
+ from peft import PeftModel
84
+ from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
85
+
86
+ repo_id = "UEC-InabaLab/CounselTraj-DA"
87
+ tokenizer = AutoTokenizer.from_pretrained(repo_id)
88
+ base = AutoModelForCausalLM.from_pretrained(
89
+ "Qwen/Qwen3-8B",
90
+ quantization_config=BitsAndBytesConfig(
91
+ load_in_4bit=True,
92
+ bnb_4bit_quant_type="nf4",
93
+ bnb_4bit_compute_dtype=torch.bfloat16,
94
+ ),
95
+ device_map="auto",
96
+ )
97
+ model = PeftModel.from_pretrained(base, repo_id)
98
+ model.eval()
99
+ ```
100
+
101
+ ## Files
102
+
103
+ - `adapter_model.safetensors` and `adapter_config.json`: selected LoRA weights and configuration.
104
+ - Tokenizer files: the saved tokenizer accompanying the adapter.
105
+ - `plan_trajectory.py`: v4 planning algorithm, with portable local defaults.
106
+ - `da_stats.json`: aggregate planning priors.
107
+ - `examples/input.jsonl`: synthetic input example.
108
+ - `model_details.json`: checkpoint selection and release details.
109
+ - `validation.json`: release checks and a synthetic GPU inference check.
110
+ - `SHA256SUMS`: file integrity checksums.
111
+
112
+ The planner generates research planning labels. Its trajectories have not been
113
+ validated as clinical treatment plans.
114
+
115
+ ## Citation
116
+
117
+ ```bibtex
118
+ @inproceedings{qi-inaba-2027-counseltraj,
119
+ title = "{CounselTraj}: Synthesizing Process-Aware Psychological Counseling Dialogues via Dialogue-Act Trajectory Planning",
120
+ author = "Qi, Zhiyang and Inaba, Michimasa",
121
+ booktitle = "Proceedings of the 20th Conference of the European Chapter of the Association for Computational Linguistics",
122
+ year = "2027",
123
+ }
124
+ ```
SHA256SUMS ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 4d49a320e87134128d77d68f82488a96b128fb1a3a9db54ea9f3f44cd289da7c README.md
2
+ d14558619c5941e26716db37de14547c202bffe9d0bffc2492bd9ea788bad2a3 adapter_config.json
3
+ f9cf6a1654d91c341678d2464098bc26c27bc58404cd57bc0a6f227d50bb6631 adapter_model.safetensors
4
+ c0284b582e14987fbd3d5a2cb2bd139084371ed9acbae488829a1c900833c680 added_tokens.json
5
+ a55ee1b1660128b7098723e0abcd92caa0788061051c62d51cbe87d9cf1974d8 chat_template.jinja
6
+ ebd1a1d94e94060f88466d87d10997daab72d235f6eb13140316ce67ca6c7a60 da_stats.json
7
+ b8955b4ed2477148b5a1b9803c4873f2bc9888498fed51e9a2d79cef8d4897b8 examples/input.jsonl
8
+ a8d98faa859e942ec71f7008abea92a21cfd7fcc3c3b26b4f02137945c502a16 examples/trajectory.json
9
+ 8831e4f1a044471340f7c0a83d7bd71306a5b867e95fd870f74d0c5308a904d5 merges.txt
10
+ 20d36d69bd1473b1c010ba32247050c062ca2f65a52f9ef21771e7866a65115a model_details.json
11
+ 07572dcf424b508c333839a33dee36263cca5af8c40826b265ae2841e60abbbc plan_trajectory.py
12
+ 5e8cce717f3c608b4496d812ff0bda2fe0419c337032a7ddd81c9869407227ed requirements.txt
13
+ 76862e765266b85aa9459767e33cbaf13970f327a0e88d1c65846c2ddd3a1ecd special_tokens_map.json
14
+ aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4 tokenizer.json
15
+ 443bfa629eb16387a12edbf92a76f6a6f10b2af3b53d87ba1550adfcf45f7fa0 tokenizer_config.json
16
+ def429f754ee9a032e5657f09b212d8bae46c7c439f9d929b9d8f7e8f58fbec6 validation.json
17
+ ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910 vocab.json
adapter_config.json ADDED
@@ -0,0 +1,42 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "alpha_pattern": {},
3
+ "auto_mapping": null,
4
+ "base_model_name_or_path": "Qwen/Qwen3-8B",
5
+ "bias": "none",
6
+ "corda_config": null,
7
+ "eva_config": null,
8
+ "exclude_modules": null,
9
+ "fan_in_fan_out": false,
10
+ "inference_mode": true,
11
+ "init_lora_weights": true,
12
+ "layer_replication": null,
13
+ "layers_pattern": null,
14
+ "layers_to_transform": null,
15
+ "loftq_config": {},
16
+ "lora_alpha": 16,
17
+ "lora_bias": false,
18
+ "lora_dropout": 0.05,
19
+ "megatron_config": null,
20
+ "megatron_core": "megatron.core",
21
+ "modules_to_save": null,
22
+ "peft_type": "LORA",
23
+ "qalora_group_size": 16,
24
+ "r": 8,
25
+ "rank_pattern": {},
26
+ "revision": null,
27
+ "target_modules": [
28
+ "v_proj",
29
+ "o_proj",
30
+ "q_proj",
31
+ "down_proj",
32
+ "up_proj",
33
+ "gate_proj",
34
+ "k_proj"
35
+ ],
36
+ "target_parameters": null,
37
+ "task_type": "CAUSAL_LM",
38
+ "trainable_token_indices": null,
39
+ "use_dora": false,
40
+ "use_qalora": false,
41
+ "use_rslora": false
42
+ }
adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f9cf6a1654d91c341678d2464098bc26c27bc58404cd57bc0a6f227d50bb6631
3
+ size 87360584
added_tokens.json ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "</think>": 151668,
3
+ "</tool_call>": 151658,
4
+ "</tool_response>": 151666,
5
+ "<think>": 151667,
6
+ "<tool_call>": 151657,
7
+ "<tool_response>": 151665,
8
+ "<|box_end|>": 151649,
9
+ "<|box_start|>": 151648,
10
+ "<|endoftext|>": 151643,
11
+ "<|file_sep|>": 151664,
12
+ "<|fim_middle|>": 151660,
13
+ "<|fim_pad|>": 151662,
14
+ "<|fim_prefix|>": 151659,
15
+ "<|fim_suffix|>": 151661,
16
+ "<|im_end|>": 151645,
17
+ "<|im_start|>": 151644,
18
+ "<|image_pad|>": 151655,
19
+ "<|object_ref_end|>": 151647,
20
+ "<|object_ref_start|>": 151646,
21
+ "<|quad_end|>": 151651,
22
+ "<|quad_start|>": 151650,
23
+ "<|repo_name|>": 151663,
24
+ "<|video_pad|>": 151656,
25
+ "<|vision_end|>": 151653,
26
+ "<|vision_pad|>": 151654,
27
+ "<|vision_start|>": 151652
28
+ }
chat_template.jinja ADDED
@@ -0,0 +1,89 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {%- if tools %}
2
+ {{- '<|im_start|>system\n' }}
3
+ {%- if messages[0].role == 'system' %}
4
+ {{- messages[0].content + '\n\n' }}
5
+ {%- endif %}
6
+ {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
7
+ {%- for tool in tools %}
8
+ {{- "\n" }}
9
+ {{- tool | tojson }}
10
+ {%- endfor %}
11
+ {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
12
+ {%- else %}
13
+ {%- if messages[0].role == 'system' %}
14
+ {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
15
+ {%- endif %}
16
+ {%- endif %}
17
+ {%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
18
+ {%- for message in messages[::-1] %}
19
+ {%- set index = (messages|length - 1) - loop.index0 %}
20
+ {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
21
+ {%- set ns.multi_step_tool = false %}
22
+ {%- set ns.last_query_index = index %}
23
+ {%- endif %}
24
+ {%- endfor %}
25
+ {%- for message in messages %}
26
+ {%- if message.content is string %}
27
+ {%- set content = message.content %}
28
+ {%- else %}
29
+ {%- set content = '' %}
30
+ {%- endif %}
31
+ {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
32
+ {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
33
+ {%- elif message.role == "assistant" %}
34
+ {%- set reasoning_content = '' %}
35
+ {%- if message.reasoning_content is string %}
36
+ {%- set reasoning_content = message.reasoning_content %}
37
+ {%- else %}
38
+ {%- if '</think>' in content %}
39
+ {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
40
+ {%- set content = content.split('</think>')[-1].lstrip('\n') %}
41
+ {%- endif %}
42
+ {%- endif %}
43
+ {%- if loop.index0 > ns.last_query_index %}
44
+ {%- if loop.last or (not loop.last and reasoning_content) %}
45
+ {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
46
+ {%- else %}
47
+ {{- '<|im_start|>' + message.role + '\n' + content }}
48
+ {%- endif %}
49
+ {%- else %}
50
+ {{- '<|im_start|>' + message.role + '\n' + content }}
51
+ {%- endif %}
52
+ {%- if message.tool_calls %}
53
+ {%- for tool_call in message.tool_calls %}
54
+ {%- if (loop.first and content) or (not loop.first) %}
55
+ {{- '\n' }}
56
+ {%- endif %}
57
+ {%- if tool_call.function %}
58
+ {%- set tool_call = tool_call.function %}
59
+ {%- endif %}
60
+ {{- '<tool_call>\n{"name": "' }}
61
+ {{- tool_call.name }}
62
+ {{- '", "arguments": ' }}
63
+ {%- if tool_call.arguments is string %}
64
+ {{- tool_call.arguments }}
65
+ {%- else %}
66
+ {{- tool_call.arguments | tojson }}
67
+ {%- endif %}
68
+ {{- '}\n</tool_call>' }}
69
+ {%- endfor %}
70
+ {%- endif %}
71
+ {{- '<|im_end|>\n' }}
72
+ {%- elif message.role == "tool" %}
73
+ {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
74
+ {{- '<|im_start|>user' }}
75
+ {%- endif %}
76
+ {{- '\n<tool_response>\n' }}
77
+ {{- content }}
78
+ {{- '\n</tool_response>' }}
79
+ {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
80
+ {{- '<|im_end|>\n' }}
81
+ {%- endif %}
82
+ {%- endif %}
83
+ {%- endfor %}
84
+ {%- if add_generation_prompt %}
85
+ {{- '<|im_start|>assistant\n' }}
86
+ {%- if enable_thinking is defined and enable_thinking is false %}
87
+ {{- '<think>\n\n</think>\n\n' }}
88
+ {%- endif %}
89
+ {%- endif %}
da_stats.json ADDED
The diff for this file is too large to render. See raw diff
 
examples/input.jsonl ADDED
@@ -0,0 +1 @@
 
 
1
+ {"id": "synthetic-example-001", "source_dataset": "synthetic_example", "source_id": "example-001", "topic": "emotion_anxiety", "topic_zh": "情绪焦虑", "input": "client:最近快考试了,我总担心考不好,晚上也睡不着。\ncounselor:听起来考试的压力让你一直很紧张。\nclient:是的,我想知道怎么让自己放松一点。"}
examples/trajectory.json ADDED
@@ -0,0 +1,49 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "index": 0,
3
+ "id": "synthetic-example-001",
4
+ "source_dataset": "synthetic_example",
5
+ "source_id": "example-001",
6
+ "topic": "emotion_anxiety",
7
+ "topic_zh": "情绪焦虑",
8
+ "turn_count": null,
9
+ "input": "client:最近快考试了,我总担心考不好,晚上也睡不着。\ncounselor:听起来考试的压力让你一直很紧张。\nclient:是的,我想知道怎么让自己放松一点。",
10
+ "features": {
11
+ "emotion": 2,
12
+ "request": 0,
13
+ "detail": 2,
14
+ "hesitation": 0,
15
+ "positive": 0,
16
+ "relation": 0,
17
+ "career": 0,
18
+ "crisis": 0,
19
+ "knowledge": 0,
20
+ "topic": "emotion_anxiety",
21
+ "topic_zh": "情绪焦虑",
22
+ "turn_count": 0,
23
+ "chars": 82
24
+ },
25
+ "topic_group": "emotion",
26
+ "target_len": 8,
27
+ "output": "counselor:[Greeting] client:[Greeting] counselor:[OpenQuestion] client:[SelfDisclosure] counselor:[Affirm] client:[DetailedDescription] counselor:[Affirm] client:[Thanking] counselor:[Greeting]",
28
+ "tag_counts": {
29
+ "Greeting": 3,
30
+ "OpenQuestion": 1,
31
+ "SelfDisclosure": 1,
32
+ "Affirm": 2,
33
+ "DetailedDescription": 1,
34
+ "Thanking": 1
35
+ },
36
+ "role_tag_counts": {
37
+ "client": {
38
+ "Greeting": 1,
39
+ "SelfDisclosure": 1,
40
+ "DetailedDescription": 1,
41
+ "Thanking": 1
42
+ },
43
+ "counselor": {
44
+ "Greeting": 2,
45
+ "OpenQuestion": 1,
46
+ "Affirm": 2
47
+ }
48
+ }
49
+ }
merges.txt ADDED
The diff for this file is too large to render. See raw diff
 
model_details.json ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "base_model": "Qwen/Qwen3-8B",
3
+ "adapter_type": "LoRA",
4
+ "selected_checkpoint": "checkpoint-145",
5
+ "epoch": 5.0,
6
+ "global_step": 145,
7
+ "validation_loss": 1.1678476333618164,
8
+ "adapter_parameters": 21823488,
9
+ "rank": 8,
10
+ "alpha": 16,
11
+ "dropout": 0.05,
12
+ "learning_rate": 0.0002,
13
+ "warmup_ratio": 0.03,
14
+ "training_epochs": 5,
15
+ "planner_version": "v4",
16
+ "prior_dialogues": 1010,
17
+ "script_changes": [
18
+ "Resolve adapter, prior and example paths relative to this release.",
19
+ "Set dialogue_chars default to 900, matching the saved production configuration."
20
+ ]
21
+ }
plan_trajectory.py ADDED
@@ -0,0 +1,929 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import argparse
2
+ import json
3
+ import math
4
+ import os
5
+ import random
6
+ from collections import Counter, defaultdict
7
+ from datetime import datetime
8
+ from pathlib import Path
9
+
10
+ import torch
11
+ import torch.nn.functional as F
12
+ from peft import PeftModel
13
+ from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
14
+
15
+
16
+ BASE_MODEL_ID = "Qwen/Qwen3-8B"
17
+ RELEASE_ROOT = Path(__file__).resolve().parent
18
+ CHECKPOINT_DIR = str(RELEASE_ROOT)
19
+ INPUT_FILE = str(RELEASE_ROOT / "examples" / "input.jsonl")
20
+ ANNOTATED_DIR = ""
21
+ OUTPUT_DIR = "./outputs"
22
+ STATS_FILE = str(RELEASE_ROOT / "da_stats.json")
23
+
24
+ CLIENT_TAGS = [
25
+ "Greeting",
26
+ "SimpleAnswer",
27
+ "DetailedDescription",
28
+ "SelfDisclosure",
29
+ "Backchannel",
30
+ "Agreement",
31
+ "PositiveEmotion",
32
+ "Hesitation",
33
+ "Question",
34
+ "Request",
35
+ "Thanking",
36
+ "UnderstandingCheck",
37
+ "Other",
38
+ ]
39
+
40
+ COUNSELOR_TAGS = [
41
+ "Greeting",
42
+ "OpenQuestion",
43
+ "ClosedQuestion",
44
+ "Paraphrase",
45
+ "Reflection",
46
+ "Affirm",
47
+ "Backchannel",
48
+ "Suggest",
49
+ "Inform",
50
+ "Thanking",
51
+ "Other",
52
+ ]
53
+
54
+ TAG_DEFINITIONS = {
55
+ "client": {
56
+ "Greeting": "opening or closing greeting",
57
+ "SimpleAnswer": "brief answer without elaboration",
58
+ "DetailedDescription": "detailed event, context, or background",
59
+ "SelfDisclosure": "inner feeling, thought, vulnerability, distress",
60
+ "Backchannel": "very short listening response",
61
+ "Agreement": "agreement with counselor's view",
62
+ "PositiveEmotion": "hope, relief, confidence, positive change",
63
+ "Hesitation": "uncertainty, difficulty speaking, ambivalence",
64
+ "Question": "general question",
65
+ "Request": "request for advice or help",
66
+ "Thanking": "thanks",
67
+ "UnderstandingCheck": "checking or restating understanding",
68
+ "Other": "other reasonable client act",
69
+ },
70
+ "counselor": {
71
+ "Greeting": "opening or closing greeting",
72
+ "OpenQuestion": "open-ended exploration question",
73
+ "ClosedQuestion": "yes/no or short factual question",
74
+ "Paraphrase": "restating facts without adding emotion",
75
+ "Reflection": "reflecting emotion or inner state",
76
+ "Affirm": "support, validation, encouragement",
77
+ "Backchannel": "short listening response",
78
+ "Suggest": "gentle advice or action option",
79
+ "Inform": "information, explanation, psychoeducation",
80
+ "Thanking": "thanks for sharing",
81
+ "Other": "other reasonable counselor act",
82
+ },
83
+ }
84
+
85
+ EMOTION_WORDS = [
86
+ "焦虑", "抑郁", "难受", "痛苦", "害怕", "恐惧", "担心", "压力", "崩溃",
87
+ "无助", "内疚", "自责", "迷茫", "纠结", "烦", "累", "失眠", "伤心",
88
+ "生气", "愤怒", "委屈", "孤独", "失败", "糟糕", "绝望",
89
+ ]
90
+ REQUEST_WORDS = ["怎么办", "建议", "帮助", "求助", "请问", "该怎么", "如何", "怎么做", "有什么办法"]
91
+ DETAIL_WORDS = ["因为", "后来", "然后", "之前", "最近", "一直", "每次", "当时", "现在", "经历"]
92
+ HESITATION_WORDS = ["不知道", "不确定", "犹豫", "纠结", "可能", "也许", "说不清", "不敢"]
93
+ POSITIVE_WORDS = ["谢谢", "感谢", "有希望", "好一点", "轻松", "明白", "愿意", "试试", "加油"]
94
+ RELATION_WORDS = ["妈妈", "爸爸", "父母", "男朋友", "女朋友", "老公", "老婆", "朋友", "同学", "老师", "孩子", "家人"]
95
+ CAREER_WORDS = ["工作", "职业", "上班", "辞职", "转行", "考研", "学习", "成绩", "学校", "专业"]
96
+ CRISIS_WORDS = [
97
+ "自杀", "自残", "轻生", "跳河", "跳楼", "割腕", "活不下去", "不想活",
98
+ "死了算了", "结束生命", "想死", "寻死", "伤害自己", "解脱",
99
+ ]
100
+ KNOWLEDGE_WORDS = [
101
+ "认知行为疗法", "疗法", "流派", "咨询师", "心理咨询", "心理治疗", "是不是",
102
+ "什么是", "区别", "原理", "理论", "方法", "技术", "诊断", "症状",
103
+ ]
104
+
105
+ EPS = 1e-8
106
+
107
+
108
+ def load_json(path):
109
+ with open(path, "r", encoding="utf-8") as f:
110
+ return json.load(f)
111
+
112
+
113
+ def save_json(path, data):
114
+ with open(path, "w", encoding="utf-8") as f:
115
+ json.dump(data, f, ensure_ascii=False, indent=2)
116
+
117
+
118
+ def read_jsonl(path, limit=None, offset=0, spread=False):
119
+ if spread and limit:
120
+ with open(path, "r", encoding="utf-8") as f:
121
+ total = sum(1 for line in f if line.strip())
122
+ if limit >= total:
123
+ wanted = set(range(total))
124
+ else:
125
+ wanted = {round(i * (total - 1) / (limit - 1)) for i in range(limit)}
126
+ rows = []
127
+ with open(path, "r", encoding="utf-8") as f:
128
+ for line_no, line in enumerate(f):
129
+ if line_no in wanted and line.strip():
130
+ obj = json.loads(line)
131
+ obj["_global_index"] = line_no
132
+ rows.append(obj)
133
+ rows.sort(key=lambda x: x["_global_index"])
134
+ return rows
135
+
136
+ rows = []
137
+ with open(path, "r", encoding="utf-8") as f:
138
+ for line_no, line in enumerate(f):
139
+ if line_no < offset:
140
+ continue
141
+ if limit is not None and len(rows) >= limit:
142
+ break
143
+ line = line.strip()
144
+ if line:
145
+ obj = json.loads(line)
146
+ obj["_global_index"] = line_no
147
+ rows.append(obj)
148
+ return rows
149
+
150
+
151
+ def append_jsonl(path, item):
152
+ with open(path, "a", encoding="utf-8") as f:
153
+ f.write(json.dumps(item, ensure_ascii=False) + "\n")
154
+
155
+
156
+ def safe_log(x):
157
+ return math.log(max(x, EPS))
158
+
159
+
160
+ def stat_key(*parts):
161
+ return "||".join(str(x) for x in parts)
162
+
163
+
164
+ def parse_stat_key(text):
165
+ return tuple(str(text).split("||"))
166
+
167
+
168
+ def get_stage(step, total):
169
+ r = step / total
170
+ if r < 0.18:
171
+ return "stage_1"
172
+ if r < 0.38:
173
+ return "stage_2"
174
+ if r < 0.62:
175
+ return "stage_3"
176
+ if r < 0.82:
177
+ return "stage_4"
178
+ return "stage_5"
179
+
180
+
181
+ def render_dialogue(item, max_chars=1200):
182
+ if item.get("input"):
183
+ text = item["input"]
184
+ else:
185
+ lines = []
186
+ for turn in item.get("dialogue", []):
187
+ role = str(turn.get("role", "")).lower()
188
+ utterance = str(turn.get("utterance", "")).strip()
189
+ if role and utterance:
190
+ lines.append(f"{role}:{utterance}")
191
+ text = "\n".join(lines)
192
+ return text[:max_chars]
193
+
194
+
195
+ def count_hits(text, words):
196
+ return sum(text.count(w) for w in words)
197
+
198
+
199
+ def text_features(item, dialogue_text):
200
+ first_client = item.get("first_client_utterance", "")
201
+ text = f"{first_client}\n{dialogue_text}"
202
+ chars = len(text)
203
+ return {
204
+ "emotion": count_hits(text, EMOTION_WORDS),
205
+ "request": count_hits(text, REQUEST_WORDS) + text.count("?") + text.count("?"),
206
+ "detail": count_hits(text, DETAIL_WORDS) + max(0, chars // 260),
207
+ "hesitation": count_hits(text, HESITATION_WORDS),
208
+ "positive": count_hits(text, POSITIVE_WORDS),
209
+ "relation": count_hits(text, RELATION_WORDS),
210
+ "career": count_hits(text, CAREER_WORDS),
211
+ "crisis": count_hits(text, CRISIS_WORDS),
212
+ "knowledge": count_hits(text, KNOWLEDGE_WORDS),
213
+ "topic": item.get("topic", ""),
214
+ "topic_zh": item.get("topic_zh", ""),
215
+ "turn_count": int(item.get("turn_count", 0) or 0),
216
+ "chars": chars,
217
+ }
218
+
219
+
220
+ def dynamic_target_len(features, min_len, max_len):
221
+ base = 58
222
+ base += min(18, features["turn_count"])
223
+ base += min(12, features["chars"] // 220)
224
+ base += min(8, features["emotion"] + features["detail"] // 2)
225
+ base += 4 if features["request"] else 0
226
+ if base % 2:
227
+ base += 1
228
+ return max(min_len, min(max_len, base))
229
+
230
+
231
+ def topic_group(features):
232
+ topic = str(features.get("topic", ""))
233
+ topic_zh = str(features.get("topic_zh", ""))
234
+ text = f"{topic} {topic_zh}".lower()
235
+
236
+ if features.get("crisis", 0):
237
+ return "crisis"
238
+ if features.get("knowledge", 0) >= 2 or any(x in text for x in ["knowledge", "therapy", "心理知识", "治疗", "咨询方式"]):
239
+ return "knowledge"
240
+ if any(x in text for x in ["academic", "school", "学业", "校园", "考试", "学习"]):
241
+ return "academic"
242
+ if any(x in text for x in ["career", "work", "职场", "职业", "工作"]):
243
+ return "career"
244
+ if any(x in text for x in ["family", "parent", "家庭", "亲子", "父母"]):
245
+ return "family"
246
+ if any(x in text for x in ["romantic", "love", "marriage", "婚恋", "恋爱", "伴侣"]):
247
+ return "romance"
248
+ if any(x in text for x in ["interpersonal", "social", "人际", "社交", "朋友", "同学"]):
249
+ return "interpersonal"
250
+ if any(x in text for x in ["depression", "anxiety", "emotion", "抑郁", "焦虑", "情绪"]):
251
+ return "emotion"
252
+ if any(x in text for x in ["health", "sleep", "body", "健康", "睡眠", "躯体"]):
253
+ return "health"
254
+ if any(x in text for x in ["habit", "behavior", "行为", "习惯"]):
255
+ return "habit"
256
+ if any(x in text for x in ["identity", "自我", "认同"]):
257
+ return "identity"
258
+ if any(x in text for x in ["growth", "planning", "成长", "规划"]):
259
+ return "planning"
260
+ if any(x in text for x in ["adaptation", "适应"]):
261
+ return "adaptation"
262
+ if any(x in text for x in ["sexual", "性心理"]):
263
+ return "sexuality"
264
+ return "general"
265
+
266
+
267
+ def topic_specific_bias(tag, role, features, step, total):
268
+ group = topic_group(features)
269
+ r = step / total
270
+ bias = 0.0
271
+
272
+ if role == "counselor":
273
+ if group == "crisis":
274
+ bias += {"ClosedQuestion": 1.10, "Reflection": 0.65, "Paraphrase": 0.35}.get(tag, 0)
275
+ bias += {"Suggest": 0.55, "Inform": 0.45}.get(tag, 0) if r > 0.32 else 0
276
+ bias += {"Backchannel": -0.35, "Thanking": -0.25}.get(tag, 0)
277
+ elif group == "knowledge":
278
+ bias += {"Inform": 1.10, "ClosedQuestion": 0.35, "OpenQuestion": 0.25, "Paraphrase": 0.20}.get(tag, 0)
279
+ bias += 0.30 if tag == "Suggest" and r > 0.55 else 0
280
+ bias += {"Reflection": -0.25, "Affirm": -0.15, "Backchannel": -0.25}.get(tag, 0)
281
+ elif group in {"academic", "career", "planning"}:
282
+ bias += {"ClosedQuestion": 0.38, "Inform": 0.58, "Suggest": 0.62, "OpenQuestion": 0.25, "Paraphrase": 0.18}.get(tag, 0)
283
+ bias += 0.18 if tag == "Reflection" and features.get("emotion", 0) else 0
284
+ bias += -0.20 if tag == "Backchannel" else 0
285
+ bias += -0.25 if tag == "Suggest" and r < 0.38 else 0
286
+ elif group in {"family", "romance", "interpersonal"}:
287
+ bias += {"Reflection": 0.78, "OpenQuestion": 0.48, "Paraphrase": 0.48, "Affirm": 0.28}.get(tag, 0)
288
+ bias += -0.35 if tag == "Inform" else 0
289
+ bias += -0.35 if tag == "Suggest" and r < 0.58 else 0
290
+ bias += 0.20 if tag == "Suggest" and r > 0.66 else 0
291
+ elif group in {"emotion", "identity"}:
292
+ bias += {"Reflection": 0.82, "Affirm": 0.35, "OpenQuestion": 0.35, "Paraphrase": 0.32}.get(tag, 0)
293
+ bias += -0.30 if tag == "Suggest" and r < 0.62 else 0
294
+ bias += -0.18 if tag == "Inform" else 0
295
+ elif group in {"health", "habit"}:
296
+ bias += {"ClosedQuestion": 0.48, "Inform": 0.45, "Suggest": 0.55, "Paraphrase": 0.25}.get(tag, 0)
297
+ bias += 0.18 if tag == "Reflection" and features.get("emotion", 0) else 0
298
+ bias += -0.20 if tag == "Suggest" and r < 0.42 else 0
299
+ elif group in {"adaptation", "sexuality"}:
300
+ bias += {"OpenQuestion": 0.40, "Reflection": 0.50, "Paraphrase": 0.35, "ClosedQuestion": 0.24}.get(tag, 0)
301
+ bias += 0.22 if tag == "Inform" and r > 0.45 else 0
302
+
303
+ else:
304
+ if group == "crisis":
305
+ bias += {"SelfDisclosure": 0.75, "Hesitation": 0.42, "DetailedDescription": 0.30, "SimpleAnswer": 0.25}.get(tag, 0)
306
+ bias += -0.95 if tag == "PositiveEmotion" and r < 0.85 else 0
307
+ bias += -0.35 if tag == "Thanking" else 0
308
+ elif group == "knowledge":
309
+ bias += {"Question": 0.85, "UnderstandingCheck": 0.75, "SimpleAnswer": 0.32, "Request": 0.28}.get(tag, 0)
310
+ bias += {"SelfDisclosure": -0.15, "Hesitation": -0.10, "PositiveEmotion": -0.25}.get(tag, 0)
311
+ elif group in {"academic", "career", "planning"}:
312
+ bias += {"DetailedDescription": 0.40, "Request": 0.35, "UnderstandingCheck": 0.30, "Agreement": 0.20}.get(tag, 0)
313
+ bias += 0.22 if tag == "SelfDisclosure" and features.get("emotion", 0) else 0
314
+ bias += -0.22 if tag == "PositiveEmotion" and r < 0.68 else 0
315
+ elif group in {"family", "romance", "interpersonal"}:
316
+ bias += {"SelfDisclosure": 0.62, "DetailedDescription": 0.46, "Hesitation": 0.32, "Agreement": 0.15}.get(tag, 0)
317
+ bias += -0.22 if tag == "Request" and r < 0.55 else 0
318
+ bias += -0.28 if tag == "PositiveEmotion" and r < 0.72 else 0
319
+ elif group in {"emotion", "identity"}:
320
+ bias += {"SelfDisclosure": 0.72, "Hesitation": 0.42, "DetailedDescription": 0.28}.get(tag, 0)
321
+ bias += -0.42 if tag == "PositiveEmotion" and r < 0.78 else 0
322
+ elif group in {"health", "habit"}:
323
+ bias += {"DetailedDescription": 0.45, "SimpleAnswer": 0.22, "Request": 0.30, "UnderstandingCheck": 0.18}.get(tag, 0)
324
+ bias += 0.18 if tag == "SelfDisclosure" and features.get("emotion", 0) else 0
325
+ elif group in {"adaptation", "sexuality"}:
326
+ bias += {"SelfDisclosure": 0.42, "Hesitation": 0.38, "DetailedDescription": 0.35, "Question": 0.22}.get(tag, 0)
327
+
328
+ return bias
329
+
330
+
331
+ def parse_tag(entry):
332
+ role, rest = entry.split(":[", 1)
333
+ return role, rest[:-1]
334
+
335
+
336
+ def tags_only(history, role=None):
337
+ out = []
338
+ for h in history:
339
+ if ":[" not in h:
340
+ continue
341
+ r, t = parse_tag(h)
342
+ if role is None or r == role:
343
+ out.append(t)
344
+ return out
345
+
346
+
347
+ def repetition_penalty(tag, history, role):
348
+ recent_all = tags_only(history[-10:])
349
+ recent_role = tags_only(history[-16:], role)
350
+ penalty = 0.0
351
+ penalty += 0.55 * recent_all.count(tag)
352
+ penalty += 0.75 * recent_role.count(tag)
353
+ if recent_role and recent_role[-1] == tag:
354
+ penalty += 2.2
355
+ if len(recent_role) >= 2 and recent_role[-2:] == [tag, tag]:
356
+ penalty += 4.0
357
+ return penalty
358
+
359
+
360
+ def diversity_bonus(tag, history, role):
361
+ recent_role = tags_only(history[-18:], role)
362
+ if tag not in recent_role:
363
+ return 1.2
364
+ if recent_role.count(tag) == 1:
365
+ return 0.35
366
+ return 0.0
367
+
368
+
369
+ def phase_bias(tag, role, step, total):
370
+ r = step / total
371
+ bias = 0.0
372
+
373
+ if role == "counselor":
374
+ if r < 0.20:
375
+ bias += {"OpenQuestion": 0.55, "Reflection": 0.35, "Thanking": 0.15}.get(tag, 0)
376
+ elif r < 0.55:
377
+ bias += {"Reflection": 0.45, "Paraphrase": 0.35, "OpenQuestion": 0.30, "Affirm": 0.20}.get(tag, 0)
378
+ elif r < 0.82:
379
+ bias += {"Suggest": 0.45, "Inform": 0.35, "Reflection": 0.25, "ClosedQuestion": 0.12}.get(tag, 0)
380
+ else:
381
+ bias += {"Suggest": 0.35, "Affirm": 0.35, "Other": 0.18}.get(tag, 0)
382
+ else:
383
+ if r < 0.18:
384
+ bias += {"DetailedDescription": 0.45, "SelfDisclosure": 0.35, "Hesitation": 0.20}.get(tag, 0)
385
+ elif r < 0.55:
386
+ bias += {"DetailedDescription": 0.35, "SelfDisclosure": 0.35, "SimpleAnswer": 0.16}.get(tag, 0)
387
+ elif r < 0.82:
388
+ bias += {"Agreement": 0.20, "Request": 0.25, "UnderstandingCheck": 0.20, "SelfDisclosure": 0.20}.get(tag, 0)
389
+ else:
390
+ bias += {"PositiveEmotion": 0.45, "Agreement": 0.25, "Thanking": 0.20}.get(tag, 0)
391
+
392
+ return bias
393
+
394
+
395
+ def content_bias(tag, role, features, step, total):
396
+ r = step / total
397
+ bias = 0.0
398
+
399
+ if role == "client":
400
+ if tag == "SelfDisclosure":
401
+ bias += min(0.85, 0.14 * features["emotion"])
402
+ if features["knowledge"] and not features["emotion"]:
403
+ bias -= 0.25
404
+ if tag == "DetailedDescription":
405
+ bias += min(0.75, 0.10 * features["detail"] + 0.04 * features["chars"] / 300)
406
+ if tag == "Request":
407
+ bias += min(0.75, 0.18 * features["request"])
408
+ if r < 0.22:
409
+ bias -= 0.25
410
+ if tag == "Question":
411
+ bias += min(0.45, 0.08 * features["request"])
412
+ bias += min(0.75, 0.16 * features["knowledge"])
413
+ if tag == "Hesitation":
414
+ bias += min(0.55, 0.14 * features["hesitation"])
415
+ if features["knowledge"]:
416
+ bias -= 0.15
417
+ if tag == "PositiveEmotion":
418
+ bias += min(0.45, 0.08 * features["positive"])
419
+ if r < 0.75:
420
+ bias -= 0.65
421
+ if features["crisis"]:
422
+ bias -= 0.80
423
+ if tag == "UnderstandingCheck":
424
+ bias += min(0.60, 0.12 * features["knowledge"])
425
+ if tag == "Thanking":
426
+ if r < 0.88:
427
+ bias -= 1.10
428
+ if features["crisis"]:
429
+ bias -= 0.35
430
+
431
+ else:
432
+ if tag == "Reflection":
433
+ bias += min(0.85, 0.12 * features["emotion"])
434
+ bias += min(0.90, 0.22 * features["crisis"])
435
+ if tag == "Paraphrase":
436
+ bias += min(0.50, 0.06 * features["detail"])
437
+ if features["knowledge"]:
438
+ bias += 0.12
439
+ if tag == "OpenQuestion":
440
+ bias += 0.25 if r < 0.45 else 0.05
441
+ if tag == "Suggest":
442
+ bias += min(0.70, 0.14 * features["request"])
443
+ if r < 0.40:
444
+ bias -= 0.35
445
+ if features["crisis"]:
446
+ bias += 0.35 if r > 0.35 else 0.05
447
+ if tag == "Inform":
448
+ bias += 0.20 if features["career"] or features["topic"] in {"academic_school", "career_work"} else 0
449
+ bias += min(1.00, 0.20 * features["knowledge"])
450
+ bias += min(0.55, 0.16 * features["crisis"])
451
+ if r < 0.35:
452
+ bias -= 0.20
453
+ if tag == "Affirm":
454
+ bias += 0.25 if features["emotion"] else 0.10
455
+ if features["knowledge"] and not features["emotion"]:
456
+ bias -= 0.10
457
+ if tag == "ClosedQuestion":
458
+ bias += min(0.90, 0.18 * features["crisis"])
459
+ if tag == "Thanking" and r > 0.35:
460
+ bias -= 0.85
461
+ if tag == "Thanking" and r < 0.88:
462
+ bias -= 1.20
463
+
464
+ if features["relation"] and role == "counselor" and tag in {"Reflection", "Paraphrase"}:
465
+ bias += 0.18
466
+
467
+ bias += topic_specific_bias(tag, role, features, step, total)
468
+
469
+ return bias
470
+
471
+
472
+ def is_allowed(tag, role, step, total, history, features):
473
+ r = step / total
474
+ if tag == "Greeting":
475
+ return False
476
+ if tag == "Thanking" and r < 0.82:
477
+ return False
478
+ if role == "counselor" and tag == "Thanking" and r < 0.88:
479
+ return False
480
+ if tag == "Backchannel":
481
+ recent_role = tags_only(history[-8:], role)
482
+ if recent_role.count("Backchannel") >= 1:
483
+ return False
484
+ if role == "counselor" and tag == "Suggest" and r < 0.30:
485
+ return False
486
+ positive_floor = 0.72 if features.get("crisis", 0) else 0.55
487
+ if role == "client" and tag == "PositiveEmotion" and r < positive_floor:
488
+ return False
489
+ return True
490
+
491
+
492
+ def build_prompt(history, dialogue_text, item, features):
493
+ tag_part = " ".join(history[-28:])
494
+ topic = item.get("topic_zh") or item.get("topic") or "unknown"
495
+ definitions = []
496
+ next_role = "client" if history[-1].startswith("counselor") else "counselor"
497
+ for tag, desc in TAG_DEFINITIONS[next_role].items():
498
+ definitions.append(f"{next_role}:[{tag}] = {desc}")
499
+ definition_text = "\n".join(definitions)
500
+
501
+ return f"""You are generating a counseling dialogue-act trajectory.
502
+ The next act must fit the short dialogue content, the topic, the phase, and the previous acts.
503
+ Prefer natural counseling flow. Avoid repetitive tag patterns.
504
+
505
+ [Topic]
506
+ {topic}
507
+
508
+ [Short Dialogue]
509
+ {dialogue_text}
510
+
511
+ [Content Signals]
512
+ emotion={features['emotion']} request={features['request']} detail={features['detail']} hesitation={features['hesitation']} crisis={features['crisis']} knowledge={features['knowledge']} topic_group={topic_group(features)}
513
+
514
+ [Recent Tag History]
515
+ {tag_part}
516
+
517
+ [Available tags for next role]
518
+ {definition_text}
519
+
520
+ Next dialogue act:
521
+ """
522
+
523
+
524
+ def load_annotated(folder):
525
+ dialogs = []
526
+ for path in Path(folder).glob("**/*.json"):
527
+ try:
528
+ obj = load_json(path)
529
+ turns = []
530
+ for turn in obj.get("dialogue", []):
531
+ role = str(turn.get("role", "")).lower()
532
+ tag = turn.get("tag") or turn.get("dialogue_tag")
533
+ if role in {"client", "counselor"} and tag:
534
+ turns.append((role, str(tag)))
535
+ if len(turns) > 3:
536
+ dialogs.append(turns)
537
+ except Exception:
538
+ continue
539
+ print(f"[INFO] loaded annotated dialogues: {len(dialogs)}", flush=True)
540
+ return dialogs
541
+
542
+
543
+ def normalize(counter):
544
+ total = sum(counter.values())
545
+ if total <= 0:
546
+ return {}
547
+ return {k: v / total for k, v in counter.items()}
548
+
549
+
550
+ def build_stats(dialogs):
551
+ stage_role = defaultdict(Counter)
552
+ transition = defaultdict(Counter)
553
+ global_role = defaultdict(Counter)
554
+ start_role = defaultdict(Counter)
555
+ end_role = defaultdict(Counter)
556
+
557
+ for dialog in dialogs:
558
+ n = len(dialog)
559
+ for i, (role, tag) in enumerate(dialog):
560
+ stage = get_stage(i, n)
561
+ stage_role[(stage, role)][tag] += 1
562
+ global_role[role][tag] += 1
563
+ if i < max(3, n // 8):
564
+ start_role[role][tag] += 1
565
+ if i >= n - max(3, n // 8):
566
+ end_role[role][tag] += 1
567
+ if i > 0:
568
+ prev_role, prev_tag = dialog[i - 1]
569
+ transition[(stage, prev_role, prev_tag, role)][tag] += 1
570
+
571
+ return {
572
+ "stage": {k: normalize(v) for k, v in stage_role.items()},
573
+ "transition": {k: normalize(v) for k, v in transition.items()},
574
+ "global": {k: normalize(v) for k, v in global_role.items()},
575
+ "start": {k: normalize(v) for k, v in start_role.items()},
576
+ "end": {k: normalize(v) for k, v in end_role.items()},
577
+ }
578
+
579
+
580
+ def load_stats_file(path):
581
+ obj = load_json(path)
582
+ raw = obj.get("stats", obj)
583
+ return {
584
+ "stage": {parse_stat_key(k): v for k, v in raw.get("stage", {}).items()},
585
+ "transition": {parse_stat_key(k): v for k, v in raw.get("transition", {}).items()},
586
+ "global": raw.get("global", {}),
587
+ "start": raw.get("start", {}),
588
+ "end": raw.get("end", {}),
589
+ }
590
+
591
+
592
+ def load_model(checkpoint_dir, use_4bit=True):
593
+ bnb_config = None
594
+ if use_4bit:
595
+ bnb_config = BitsAndBytesConfig(
596
+ load_in_4bit=True,
597
+ bnb_4bit_quant_type="nf4",
598
+ bnb_4bit_compute_dtype=torch.bfloat16,
599
+ )
600
+
601
+ tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL_ID, trust_remote_code=True)
602
+ if tokenizer.pad_token is None:
603
+ tokenizer.pad_token = tokenizer.eos_token
604
+
605
+ base_model = AutoModelForCausalLM.from_pretrained(
606
+ BASE_MODEL_ID,
607
+ quantization_config=bnb_config,
608
+ device_map="auto",
609
+ trust_remote_code=True,
610
+ )
611
+ model = PeftModel.from_pretrained(base_model, checkpoint_dir)
612
+ model.eval()
613
+ return tokenizer, model
614
+
615
+
616
+ def batch_conditional_lm_scores(model, tokenizer, prompt, candidates, max_length):
617
+ prefix_ids = tokenizer(prompt, add_special_tokens=False)["input_ids"]
618
+ texts = [prompt + " " + candidate for candidate in candidates]
619
+ enc = tokenizer(
620
+ texts,
621
+ return_tensors="pt",
622
+ add_special_tokens=False,
623
+ truncation=True,
624
+ max_length=max_length,
625
+ padding=True,
626
+ )
627
+ input_ids = enc["input_ids"].to(model.device)
628
+ attention_mask = enc["attention_mask"].to(model.device)
629
+
630
+ with torch.inference_mode():
631
+ outputs = model(input_ids=input_ids, attention_mask=attention_mask)
632
+
633
+ logits = outputs.logits[:, :-1, :]
634
+ labels = input_ids[:, 1:]
635
+ mask = attention_mask[:, 1:].float()
636
+ log_probs = F.log_softmax(logits, dim=-1)
637
+ selected = log_probs.gather(2, labels.unsqueeze(-1)).squeeze(-1)
638
+
639
+ start = max(0, min(selected.shape[1] - 1, len(prefix_ids) - 1))
640
+ candidate_mask = torch.zeros_like(mask)
641
+ candidate_mask[:, start:] = mask[:, start:]
642
+ scores = (selected * candidate_mask).sum(dim=1) / candidate_mask.sum(dim=1).clamp_min(1.0)
643
+
644
+ out = scores.detach().cpu().tolist()
645
+ del outputs
646
+ return out
647
+
648
+
649
+ def candidate_tags(role, step, total, history, stats, features):
650
+ stage = get_stage(step, total)
651
+ base = COUNSELOR_TAGS if role == "counselor" else CLIENT_TAGS
652
+ stat_tags = set(stats["stage"].get((stage, role), {}).keys())
653
+ stat_tags.update(stats["global"].get(role, {}).keys())
654
+ tags = [tag for tag in base if tag in stat_tags or tag in base]
655
+ return [tag for tag in tags if is_allowed(tag, role, step, total, history, features)]
656
+
657
+
658
+ def choose_candidate(scored, rng, temperature, top_k, top_p, deterministic_gap):
659
+ scored = sorted(scored, key=lambda x: x["score"], reverse=True)
660
+ if not scored:
661
+ return None, []
662
+
663
+ gap = scored[0]["score"] - scored[1]["score"] if len(scored) > 1 else float("inf")
664
+ if gap > deterministic_gap:
665
+ return scored[0], scored[:top_k]
666
+
667
+ pool = scored[: max(1, min(top_k, len(scored)))]
668
+ raw = torch.tensor([x["score"] for x in pool], dtype=torch.float32)
669
+ probs = torch.softmax(raw / max(temperature, 1e-4), dim=0).tolist()
670
+
671
+ ranked = sorted(zip(pool, probs), key=lambda x: x[1], reverse=True)
672
+ kept = []
673
+ acc = 0.0
674
+ for item, prob in ranked:
675
+ kept.append((item, prob))
676
+ acc += prob
677
+ if acc >= top_p:
678
+ break
679
+
680
+ total = sum(prob for _, prob in kept)
681
+ pick = rng.random() * total
682
+ acc = 0.0
683
+ for item, prob in kept:
684
+ acc += prob
685
+ if pick <= acc:
686
+ return item, pool
687
+ return kept[-1][0], pool
688
+
689
+
690
+ def finalize_history(history):
691
+ last_role, _ = parse_tag(history[-1])
692
+ if last_role == "counselor":
693
+ history.append("client:[Thanking]")
694
+ history.append("counselor:[Greeting]")
695
+ else:
696
+ history.append("counselor:[Affirm]")
697
+ history.append("client:[Thanking]")
698
+ history.append("counselor:[Greeting]")
699
+ return history
700
+
701
+
702
+ def generate_trajectory(item, tokenizer, model, stats, args, sample_index):
703
+ dialogue_text = render_dialogue(item, max_chars=args.dialogue_chars)
704
+ features = text_features(item, dialogue_text)
705
+ target_len = dynamic_target_len(features, args.min_len, args.max_len)
706
+
707
+ rng_seed = args.seed + sample_index * 1009 + sum(ord(c) for c in str(item.get("id", "")))
708
+ rng = random.Random(rng_seed)
709
+
710
+ history = ["counselor:[Greeting]", "client:[Greeting]"]
711
+ role = "counselor"
712
+ trace = []
713
+
714
+ for step in range(2, target_len - 2):
715
+ stage = get_stage(step, target_len)
716
+ prompt = build_prompt(history, dialogue_text, item, features)
717
+
718
+ prev_role, prev_tag = parse_tag(history[-1])
719
+ tags = candidate_tags(role, step, target_len, history, stats, features)
720
+ if not tags:
721
+ tags = ["Other"]
722
+
723
+ candidates = [f"{role}:[{tag}]" for tag in tags]
724
+ lm_scores = batch_conditional_lm_scores(model, tokenizer, prompt, candidates, args.max_length)
725
+
726
+ scored = []
727
+ for tag, candidate, lm in zip(tags, candidates, lm_scores):
728
+ s = safe_log(stats["stage"].get((stage, role), {}).get(tag, 1e-6))
729
+ g = safe_log(stats["global"].get(role, {}).get(tag, 1e-6))
730
+ t = safe_log(stats["transition"].get((stage, prev_role, prev_tag, role), {}).get(tag, 1e-6))
731
+ rep = repetition_penalty(tag, history, role)
732
+ div = diversity_bonus(tag, history, role)
733
+ cb = content_bias(tag, role, features, step, target_len)
734
+ pb = phase_bias(tag, role, step, target_len)
735
+
736
+ score = (
737
+ args.w_lm * lm
738
+ + args.w_stage * s
739
+ + args.w_trans * t
740
+ + args.w_global * g
741
+ - args.w_repeat * rep
742
+ + args.w_diversity * div
743
+ + args.w_content * cb
744
+ + args.w_phase * pb
745
+ )
746
+ scored.append({
747
+ "candidate": candidate,
748
+ "tag": tag,
749
+ "score": score,
750
+ "lm": lm,
751
+ "content": cb,
752
+ "phase": pb,
753
+ "repeat": rep,
754
+ })
755
+
756
+ best, pool = choose_candidate(
757
+ scored,
758
+ rng=rng,
759
+ temperature=args.temperature,
760
+ top_k=args.top_k,
761
+ top_p=args.top_p,
762
+ deterministic_gap=args.deterministic_gap,
763
+ )
764
+
765
+ history.append(best["candidate"])
766
+ if args.keep_trace:
767
+ trace.append({
768
+ "step": step,
769
+ "role": role,
770
+ "chosen": best,
771
+ "top": pool[: min(5, len(pool))],
772
+ })
773
+
774
+ role = "client" if role == "counselor" else "counselor"
775
+
776
+ history = finalize_history(history)
777
+
778
+ tag_counts = Counter(tags_only(history))
779
+ role_counts = {
780
+ "client": Counter(tags_only(history, "client")),
781
+ "counselor": Counter(tags_only(history, "counselor")),
782
+ }
783
+
784
+ result = {
785
+ "index": sample_index,
786
+ "id": item.get("id", item.get("dialogue_id", str(sample_index))),
787
+ "source_dataset": item.get("source_dataset", ""),
788
+ "source_id": item.get("source_id", ""),
789
+ "topic": item.get("topic", ""),
790
+ "topic_zh": item.get("topic_zh", ""),
791
+ "turn_count": item.get("turn_count", None),
792
+ "input": dialogue_text,
793
+ "features": features,
794
+ "topic_group": topic_group(features),
795
+ "target_len": target_len,
796
+ "output": " ".join(history),
797
+ "tag_counts": dict(tag_counts),
798
+ "role_tag_counts": {
799
+ "client": dict(role_counts["client"]),
800
+ "counselor": dict(role_counts["counselor"]),
801
+ },
802
+ }
803
+ if args.keep_trace:
804
+ result["trace"] = trace
805
+ return result
806
+
807
+
808
+ def output_path(args):
809
+ name = args.output_name
810
+ if not name:
811
+ stamp = datetime.now().strftime("%Y%m%d_%H%M%S")
812
+ count = "all" if args.limit is None else str(args.limit)
813
+ name = f"da_trajectories_offset{args.offset}_n{count}_{stamp}.jsonl"
814
+ return Path(args.output_dir) / name
815
+
816
+
817
+ def parse_args():
818
+ parser = argparse.ArgumentParser()
819
+ parser.add_argument("--input-file", default=INPUT_FILE)
820
+ parser.add_argument("--annotated-dir", default=ANNOTATED_DIR)
821
+ parser.add_argument("--checkpoint-dir", default=CHECKPOINT_DIR)
822
+ parser.add_argument("--output-dir", default=OUTPUT_DIR)
823
+ parser.add_argument("--output-name", default="")
824
+ parser.add_argument("--stats-file", default=STATS_FILE)
825
+ parser.add_argument("--offset", type=int, default=0)
826
+ parser.add_argument("--limit", type=int, default=20)
827
+ parser.add_argument("--spread", action="store_true")
828
+ parser.add_argument("--num-shards", type=int, default=1)
829
+ parser.add_argument("--shard-id", type=int, default=0)
830
+ parser.add_argument("--seed", type=int, default=20260703)
831
+ parser.add_argument("--min-len", type=int, default=64)
832
+ parser.add_argument("--max-len", type=int, default=92)
833
+ parser.add_argument("--dialogue-chars", type=int, default=900)
834
+ parser.add_argument("--max-length", type=int, default=2048)
835
+ parser.add_argument("--top-k", type=int, default=5)
836
+ parser.add_argument("--top-p", type=float, default=0.88)
837
+ parser.add_argument("--temperature", type=float, default=0.64)
838
+ parser.add_argument("--deterministic-gap", type=float, default=0.48)
839
+ parser.add_argument("--w-lm", type=float, default=4.0)
840
+ parser.add_argument("--w-stage", type=float, default=0.08)
841
+ parser.add_argument("--w-trans", type=float, default=0.12)
842
+ parser.add_argument("--w-global", type=float, default=0.01)
843
+ parser.add_argument("--w-repeat", type=float, default=1.85)
844
+ parser.add_argument("--w-diversity", type=float, default=0.52)
845
+ parser.add_argument("--w-content", type=float, default=1.18)
846
+ parser.add_argument("--w-phase", type=float, default=0.50)
847
+ parser.add_argument("--no-4bit", action="store_true")
848
+ parser.add_argument("--keep-trace", action="store_true")
849
+ parser.add_argument("--resume", action="store_true")
850
+ return parser.parse_args()
851
+
852
+
853
+ def main():
854
+ args = parse_args()
855
+ if args.limit is not None and args.limit <= 0:
856
+ args.limit = None
857
+ os.makedirs(args.output_dir, exist_ok=True)
858
+ out_path = output_path(args)
859
+ summary_path = out_path.with_suffix(".summary.json")
860
+ config_path = out_path.with_suffix(".config.json")
861
+
862
+ save_json(config_path, vars(args))
863
+
864
+ done_ids = set()
865
+ if args.resume and out_path.exists():
866
+ with open(out_path, "r", encoding="utf-8") as f:
867
+ for line in f:
868
+ if line.strip():
869
+ done_ids.add(json.loads(line).get("id"))
870
+ print(f"[INFO] resume enabled, loaded done ids: {len(done_ids)}", flush=True)
871
+ elif out_path.exists():
872
+ out_path.unlink()
873
+
874
+ data = read_jsonl(args.input_file, limit=args.limit, offset=args.offset, spread=args.spread)
875
+ if args.num_shards > 1:
876
+ if args.shard_id < 0 or args.shard_id >= args.num_shards:
877
+ raise ValueError("--shard-id must be in [0, --num-shards)")
878
+ data = [item for pos, item in enumerate(data) if pos % args.num_shards == args.shard_id]
879
+ print(f"[INFO] loaded input rows: {len(data)}", flush=True)
880
+ print(f"[INFO] output: {out_path}", flush=True)
881
+
882
+ tokenizer, model = load_model(args.checkpoint_dir, use_4bit=not args.no_4bit)
883
+ if args.stats_file and os.path.exists(args.stats_file):
884
+ stats = load_stats_file(args.stats_file)
885
+ print(f"[INFO] loaded stats file: {args.stats_file}", flush=True)
886
+ else:
887
+ stats = build_stats(load_annotated(args.annotated_dir))
888
+
889
+ generated = []
890
+ for local_i, item in enumerate(data):
891
+ sample_index = int(item.get("_global_index", args.offset + local_i))
892
+ item_id = item.get("id", str(sample_index))
893
+ if item_id in done_ids:
894
+ continue
895
+
896
+ print("\n" + "-" * 60, flush=True)
897
+ print(f"[INFO] sample {local_i + 1}/{len(data)} global_index={sample_index} id={item_id}", flush=True)
898
+ result = generate_trajectory(item, tokenizer, model, stats, args, sample_index)
899
+ append_jsonl(out_path, result)
900
+ generated.append(result)
901
+
902
+ print(
903
+ f"[OK] len={result['target_len']} topic={result.get('topic_zh') or result.get('topic')} "
904
+ f"distinct_tags={len(result['tag_counts'])}",
905
+ flush=True,
906
+ )
907
+ print(result["output"], flush=True)
908
+
909
+ lengths = [x["target_len"] for x in generated]
910
+ all_tags = Counter()
911
+ for x in generated:
912
+ all_tags.update(x["tag_counts"])
913
+
914
+ summary = {
915
+ "output": str(out_path),
916
+ "generated": len(generated),
917
+ "offset": args.offset,
918
+ "limit": args.limit,
919
+ "length_min": min(lengths) if lengths else None,
920
+ "length_max": max(lengths) if lengths else None,
921
+ "length_avg": sum(lengths) / len(lengths) if lengths else None,
922
+ "tag_counts": dict(all_tags),
923
+ }
924
+ save_json(summary_path, summary)
925
+ print("\n[INFO] summary:", json.dumps(summary, ensure_ascii=False, indent=2), flush=True)
926
+
927
+
928
+ if __name__ == "__main__":
929
+ main()
requirements.txt ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ torch>=2.5
2
+ transformers>=4.57.6,<5
3
+ peft>=0.17,<1
4
+ accelerate>=1
5
+ bitsandbytes>=0.45
6
+ huggingface_hub>=0.34
7
+ safetensors>=0.5
special_tokens_map.json ADDED
@@ -0,0 +1,31 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "additional_special_tokens": [
3
+ "<|im_start|>",
4
+ "<|im_end|>",
5
+ "<|object_ref_start|>",
6
+ "<|object_ref_end|>",
7
+ "<|box_start|>",
8
+ "<|box_end|>",
9
+ "<|quad_start|>",
10
+ "<|quad_end|>",
11
+ "<|vision_start|>",
12
+ "<|vision_end|>",
13
+ "<|vision_pad|>",
14
+ "<|image_pad|>",
15
+ "<|video_pad|>"
16
+ ],
17
+ "eos_token": {
18
+ "content": "<|im_end|>",
19
+ "lstrip": false,
20
+ "normalized": false,
21
+ "rstrip": false,
22
+ "single_word": false
23
+ },
24
+ "pad_token": {
25
+ "content": "<|endoftext|>",
26
+ "lstrip": false,
27
+ "normalized": false,
28
+ "rstrip": false,
29
+ "single_word": false
30
+ }
31
+ }
tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4
3
+ size 11422654
tokenizer_config.json ADDED
@@ -0,0 +1,239 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_bos_token": false,
3
+ "add_prefix_space": false,
4
+ "added_tokens_decoder": {
5
+ "151643": {
6
+ "content": "<|endoftext|>",
7
+ "lstrip": false,
8
+ "normalized": false,
9
+ "rstrip": false,
10
+ "single_word": false,
11
+ "special": true
12
+ },
13
+ "151644": {
14
+ "content": "<|im_start|>",
15
+ "lstrip": false,
16
+ "normalized": false,
17
+ "rstrip": false,
18
+ "single_word": false,
19
+ "special": true
20
+ },
21
+ "151645": {
22
+ "content": "<|im_end|>",
23
+ "lstrip": false,
24
+ "normalized": false,
25
+ "rstrip": false,
26
+ "single_word": false,
27
+ "special": true
28
+ },
29
+ "151646": {
30
+ "content": "<|object_ref_start|>",
31
+ "lstrip": false,
32
+ "normalized": false,
33
+ "rstrip": false,
34
+ "single_word": false,
35
+ "special": true
36
+ },
37
+ "151647": {
38
+ "content": "<|object_ref_end|>",
39
+ "lstrip": false,
40
+ "normalized": false,
41
+ "rstrip": false,
42
+ "single_word": false,
43
+ "special": true
44
+ },
45
+ "151648": {
46
+ "content": "<|box_start|>",
47
+ "lstrip": false,
48
+ "normalized": false,
49
+ "rstrip": false,
50
+ "single_word": false,
51
+ "special": true
52
+ },
53
+ "151649": {
54
+ "content": "<|box_end|>",
55
+ "lstrip": false,
56
+ "normalized": false,
57
+ "rstrip": false,
58
+ "single_word": false,
59
+ "special": true
60
+ },
61
+ "151650": {
62
+ "content": "<|quad_start|>",
63
+ "lstrip": false,
64
+ "normalized": false,
65
+ "rstrip": false,
66
+ "single_word": false,
67
+ "special": true
68
+ },
69
+ "151651": {
70
+ "content": "<|quad_end|>",
71
+ "lstrip": false,
72
+ "normalized": false,
73
+ "rstrip": false,
74
+ "single_word": false,
75
+ "special": true
76
+ },
77
+ "151652": {
78
+ "content": "<|vision_start|>",
79
+ "lstrip": false,
80
+ "normalized": false,
81
+ "rstrip": false,
82
+ "single_word": false,
83
+ "special": true
84
+ },
85
+ "151653": {
86
+ "content": "<|vision_end|>",
87
+ "lstrip": false,
88
+ "normalized": false,
89
+ "rstrip": false,
90
+ "single_word": false,
91
+ "special": true
92
+ },
93
+ "151654": {
94
+ "content": "<|vision_pad|>",
95
+ "lstrip": false,
96
+ "normalized": false,
97
+ "rstrip": false,
98
+ "single_word": false,
99
+ "special": true
100
+ },
101
+ "151655": {
102
+ "content": "<|image_pad|>",
103
+ "lstrip": false,
104
+ "normalized": false,
105
+ "rstrip": false,
106
+ "single_word": false,
107
+ "special": true
108
+ },
109
+ "151656": {
110
+ "content": "<|video_pad|>",
111
+ "lstrip": false,
112
+ "normalized": false,
113
+ "rstrip": false,
114
+ "single_word": false,
115
+ "special": true
116
+ },
117
+ "151657": {
118
+ "content": "<tool_call>",
119
+ "lstrip": false,
120
+ "normalized": false,
121
+ "rstrip": false,
122
+ "single_word": false,
123
+ "special": false
124
+ },
125
+ "151658": {
126
+ "content": "</tool_call>",
127
+ "lstrip": false,
128
+ "normalized": false,
129
+ "rstrip": false,
130
+ "single_word": false,
131
+ "special": false
132
+ },
133
+ "151659": {
134
+ "content": "<|fim_prefix|>",
135
+ "lstrip": false,
136
+ "normalized": false,
137
+ "rstrip": false,
138
+ "single_word": false,
139
+ "special": false
140
+ },
141
+ "151660": {
142
+ "content": "<|fim_middle|>",
143
+ "lstrip": false,
144
+ "normalized": false,
145
+ "rstrip": false,
146
+ "single_word": false,
147
+ "special": false
148
+ },
149
+ "151661": {
150
+ "content": "<|fim_suffix|>",
151
+ "lstrip": false,
152
+ "normalized": false,
153
+ "rstrip": false,
154
+ "single_word": false,
155
+ "special": false
156
+ },
157
+ "151662": {
158
+ "content": "<|fim_pad|>",
159
+ "lstrip": false,
160
+ "normalized": false,
161
+ "rstrip": false,
162
+ "single_word": false,
163
+ "special": false
164
+ },
165
+ "151663": {
166
+ "content": "<|repo_name|>",
167
+ "lstrip": false,
168
+ "normalized": false,
169
+ "rstrip": false,
170
+ "single_word": false,
171
+ "special": false
172
+ },
173
+ "151664": {
174
+ "content": "<|file_sep|>",
175
+ "lstrip": false,
176
+ "normalized": false,
177
+ "rstrip": false,
178
+ "single_word": false,
179
+ "special": false
180
+ },
181
+ "151665": {
182
+ "content": "<tool_response>",
183
+ "lstrip": false,
184
+ "normalized": false,
185
+ "rstrip": false,
186
+ "single_word": false,
187
+ "special": false
188
+ },
189
+ "151666": {
190
+ "content": "</tool_response>",
191
+ "lstrip": false,
192
+ "normalized": false,
193
+ "rstrip": false,
194
+ "single_word": false,
195
+ "special": false
196
+ },
197
+ "151667": {
198
+ "content": "<think>",
199
+ "lstrip": false,
200
+ "normalized": false,
201
+ "rstrip": false,
202
+ "single_word": false,
203
+ "special": false
204
+ },
205
+ "151668": {
206
+ "content": "</think>",
207
+ "lstrip": false,
208
+ "normalized": false,
209
+ "rstrip": false,
210
+ "single_word": false,
211
+ "special": false
212
+ }
213
+ },
214
+ "additional_special_tokens": [
215
+ "<|im_start|>",
216
+ "<|im_end|>",
217
+ "<|object_ref_start|>",
218
+ "<|object_ref_end|>",
219
+ "<|box_start|>",
220
+ "<|box_end|>",
221
+ "<|quad_start|>",
222
+ "<|quad_end|>",
223
+ "<|vision_start|>",
224
+ "<|vision_end|>",
225
+ "<|vision_pad|>",
226
+ "<|image_pad|>",
227
+ "<|video_pad|>"
228
+ ],
229
+ "bos_token": null,
230
+ "clean_up_tokenization_spaces": false,
231
+ "eos_token": "<|im_end|>",
232
+ "errors": "replace",
233
+ "extra_special_tokens": {},
234
+ "model_max_length": 131072,
235
+ "pad_token": "<|endoftext|>",
236
+ "split_special_tokens": false,
237
+ "tokenizer_class": "Qwen2Tokenizer",
238
+ "unk_token": null
239
+ }
validation.json ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "adapter_sha256": "f9cf6a1654d91c341678d2464098bc26c27bc58404cd57bc0a6f227d50bb6631",
3
+ "matches_selected_checkpoint_and_final_model": true,
4
+ "adapter_tensors": 504,
5
+ "adapter_parameters": 21823488,
6
+ "all_tensors_finite": true,
7
+ "model_card_and_python_syntax_validated": true,
8
+ "planning_functions_match_original": true,
9
+ "gpu_smoke_test": {
10
+ "passed": true,
11
+ "input": "examples/input.jsonl",
12
+ "output": "examples/trajectory.json",
13
+ "target_length_budget": 8,
14
+ "realized_acts": 9,
15
+ "max_input_length": 512,
16
+ "quantization": "NF4",
17
+ "torch": "2.5.1+cu121",
18
+ "transformers": "4.57.6",
19
+ "peft": "0.17.1",
20
+ "bitsandbytes": "0.49.2"
21
+ }
22
+ }
vocab.json ADDED
The diff for this file is too large to render. See raw diff