Hugging Face's logo

Datasets:
SeanWang0027
/
sciworld-teacher32b-sft-partial

Modalities:
Tabular
Text
Formats:
json
Size:
< 1K
Libraries:
Datasets
Dask
Polars
Dataset card Data Studio Files
xet
 Community
1
Dataset Viewer
Auto-converted to Parquet Duplicate
messages
listlengths
3
61
loss_mask
listlengths
3
61
score
float64
0
1
task_desc
stringlengths
65
109
n_turns
int64
1
30
n_noaction
int64
0
14
done
bool
2 classes
[ { "role": "system", "content": "You are an embodied agent solving a science task in the ScienceWorld text-based simulator. You receive an Observation describing what you can see and the action commands and objects available to you. On every turn you must reply with exactly two lines:\nThought: <one short se...
[ 0, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1 ]
0.16
{"task_name": "mendelian-genetics-unknown-plant", "var_num": 53, "simplification_str": "easy"}
9
3
true
[ { "role": "system", "content": "You are an embodied agent solving a science task in the ScienceWorld text-based simulator. You receive an Observation describing what you can see and the action commands and objects available to you. On every turn you must reply with exactly two lines:\nThought: <one short se...
[ 0, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1 ]
1
{"task_name": "mendelian-genetics-unknown-plant", "var_num": 36, "simplification_str": "easy"}
21
7
true
[{"role":"system","content":"You are an embodied agent solving a science task in the ScienceWorld te(...TRUNCATED)
[0,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0(...TRUNCATED)
0.74
"{\"task_name\": \"test-conductivity-of-unknown-substances\", \"var_num\": 289, \"simplification_str(...TRUNCATED)
30
10
false
[{"role":"system","content":"You are an embodied agent solving a science task in the ScienceWorld te(...TRUNCATED)
[0,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0(...TRUNCATED)
0.17
{"task_name": "find-animal", "var_num": 126, "simplification_str": "easy"}
30
0
false
[{"role":"system","content":"You are an embodied agent solving a science task in the ScienceWorld te(...TRUNCATED)
[ 0, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1 ]
1
{"task_name": "find-animal", "var_num": 94, "simplification_str": "easy"}
9
1
true
[{"role":"system","content":"You are an embodied agent solving a science task in the ScienceWorld te(...TRUNCATED)
[ 0, 0, 1, 0, 1, 0, 1, 0, 1 ]
1
{"task_name": "lifespan-longest-lived", "var_num": 31, "simplification_str": "easy"}
4
0
true
[{"role":"system","content":"You are an embodied agent solving a science task in the ScienceWorld te(...TRUNCATED)
[0,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0(...TRUNCATED)
0.4
"{\"task_name\": \"inclined-plane-friction-named-surfaces\", \"var_num\": 410, \"simplification_str\(...TRUNCATED)
30
4
false
[{"role":"system","content":"You are an embodied agent solving a science task in the ScienceWorld te(...TRUNCATED)
[ 0, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1 ]
0.15
"{\"task_name\": \"inclined-plane-friction-unnamed-surfaces\", \"var_num\": 34, \"simplification_str(...TRUNCATED)
19
7
true
[{"role":"system","content":"You are an embodied agent solving a science task in the ScienceWorld te(...TRUNCATED)
[0,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0(...TRUNCATED)
1
{"task_name": "find-non-living-thing", "var_num": 148, "simplification_str": "easy"}
27
3
true
[{"role":"system","content":"You are an embodied agent solving a science task in the ScienceWorld te(...TRUNCATED)
[0,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0(...TRUNCATED)
0.03
"{\"task_name\": \"change-the-state-of-matter-of\", \"var_num\": 6, \"simplification_str\": \"easy\"(...TRUNCATED)
30
2
false
End of preview. Expand in Data Studio

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

SciWorld — Qwen3-32B teacher trajectories for SFT (partial)

用 Qwen3-32B 在 ScienceWorld 训练集上跑完整轨迹, 导出成 SFT 数据。 这是一批不完整的采集: 目标 3592 条, 实际只成功 158 条。

⚠ 为什么不完整

采集脚本以 48 并发跑, 每条轨迹独占一个 ScienceWorld env(= 一个 JVM + py4j socket)。默认 fd 上限 1024 被耗尽, 3434 条挂在 OSError(24, Too many open files)。 scripts/run_collect.sh 已修(加 ulimit -n 65536 与失败率闸门), 但这批数据是 修复前采的。当作抽样看, 不要当作完整训练集。

数据

messages + loss_mask 格式, 与 TCOD_examples/rose/rose_export_sft.py 对齐:

{"messages": [{"role":"system"|"user"|"assistant","content":...}, ...],
 "loss_mask": [0,0,1,0,1,...],
 "score": 0.0-1.0, "task_desc": ..., "n_turns": ..., "done": bool}

loss_mask 与 messages 对齐, 只有 assistant turn 是 1。

文件 轨迹 说明
sciworld_sft_teacher32b_all.jsonl 158 全部, 未筛
sciworld_sft_teacher32b_s100.jsonl 40 score = 1.0
sciworld_sft_teacher32b_s080.jsonl 46 score >= 0.8
sciworld_sft_teacher32b_s050.jsonl 108 score >= 0.5
  • 平均 21.5 turn/轨迹, 共 3404 个可训 turn
  • 平均 score 0.6086, 完全做对 40/158 (25.3%)
  • score 分布: {0.0: 3, 0.1: 15, 0.2: 15, 0.3: 6, 0.4: 11, 0.5: 4, 0.6: 28, 0.7: 12, 0.8: 24, 1.0: 40}

⚠ assistant 内容带完整 <think> 段

3404/3404 个可训 turn 的内容以 <think> 开头 —— 32B 先写一大段 thinking, 再写 Thought: + <action>。训练前需要决定:

  • 原样训 — 学生要学完整的 think + Thought + action
  • 剥掉 <think>...</think> — 只留 Thought: + <action>, 目标短得多, 也更接近 SAGE-OPD 的原意。1.7B 学不动 32B 的长 thinking(在 RLVE 上反复验证过), 且推理时 这些 thinking 会让多轮任务的每步成本翻几倍。

prompt style

采集时 SCIWORLD_PROMPT_STYLE=thought(SAGE-OPD 格式)。默认的 think style 在 Qwen3 chat template 跑 enable_thinking=False 时是失效的 —— 模板在模型开写前就 自动输出并闭合了空的 <think></think>, 实测 11k turn 里 <think> 出现率 0.0%。

复现 

export SCIWORLD_PROMPT_STYLE=thought
export JAVA_HOME=/path/to/jre        # ScienceWorldEnv 靠 py4j 拉 JVM, 缺了就 FileNotFoundError
export PATH=$JAVA_HOME/bin:$PATH
ulimit -n 65536                       # 48 并发下必须, 否则 Too many open files
python3 scripts/collect_teacher_sft.py \
  --tasks data/sciworld/train.jsonl --out out.jsonl \
  --teacher-url http://127.0.0.1:8000/v1 --teacher-model teacher \
  --max-env-steps 30 --concurrency 48 --keep-all

参考: SciWorld 上已有的分数 (thought style)

test 1819 mean_score success
1.7B base 0.0534 0.0077
OPD s250 0.3238 0.1759
ROSE s250 0.3176 0.1594
32B teacher 0.3733 0.2056
gold(脚本最优) 0.7048 0.4450

这批采集的平均 score 0.6086 高于 32B 在 test 上的 0.3733, 但样本只有 158 条且因失败而有选择偏差(能跑完的任务可能系统性更简单), 不要直接对比。
Downloads last month
64
Number of rows:
352
Total file size:
31.5 MB