Dataset Viewer
Auto-converted to Parquet Duplicate
messages
listlengths
2
130
tools
stringclasses
1 value
source
stringclasses
1 value
metadata
dict
[ { "content": "You are a helpful assistant that can interact with a computer.\n\nYour response must include a THOUGHT section before your action where you\nexplain your reasoning. After the THOUGHT, you must call the `bash` tool\nwith EXACTLY ONE bash command (multiple commands chained with `&&` or `||`\ncount a...
[{"type": "function", "function": {"name": "bash", "description": "Execute a bash command. Each command runs in a new subshell.", "parameters": {"type": "object", "properties": {"command": {"type": "string", "description": "The bash command to execute."}}, "required": ["command"]}}}]
deepseek-v4-flash-official
{ "task_id": "task_000001_27b0aa02", "domain": "debugging", "reward": 1, "k_index": 0, "n_bash": 5, "steps_taken": 5, "model": "deepseek-v4-flash" }
[ { "content": "You are a helpful assistant that can interact with a computer.\n\nYour response must include a THOUGHT section before your action where you\nexplain your reasoning. After the THOUGHT, you must call the `bash` tool\nwith EXACTLY ONE bash command (multiple commands chained with `&&` or `||`\ncount a...
[{"type": "function", "function": {"name": "bash", "description": "Execute a bash command. Each command runs in a new subshell.", "parameters": {"type": "object", "properties": {"command": {"type": "string", "description": "The bash command to execute."}}, "required": ["command"]}}}]
deepseek-v4-flash-official
{ "task_id": "task_000012_e76205b0", "domain": "data_processing", "reward": 1, "k_index": 0, "n_bash": 6, "steps_taken": 6, "model": "deepseek-v4-flash" }
[ { "content": "You are a helpful assistant that can interact with a computer.\n\nYour response must include a THOUGHT section before your action where you\nexplain your reasoning. After the THOUGHT, you must call the `bash` tool\nwith EXACTLY ONE bash command (multiple commands chained with `&&` or `||`\ncount a...
[{"type": "function", "function": {"name": "bash", "description": "Execute a bash command. Each command runs in a new subshell.", "parameters": {"type": "object", "properties": {"command": {"type": "string", "description": "The bash command to execute."}}, "required": ["command"]}}}]
deepseek-v4-flash-official
{ "task_id": "task_000000_5c5b0938", "domain": "scientific_computing", "reward": 1, "k_index": 0, "n_bash": 6, "steps_taken": 6, "model": "deepseek-v4-flash" }
[ { "content": "You are a helpful assistant that can interact with a computer.\n\nYour response must include a THOUGHT section before your action where you\nexplain your reasoning. After the THOUGHT, you must call the `bash` tool\nwith EXACTLY ONE bash command (multiple commands chained with `&&` or `||`\ncount a...
[{"type": "function", "function": {"name": "bash", "description": "Execute a bash command. Each command runs in a new subshell.", "parameters": {"type": "object", "properties": {"command": {"type": "string", "description": "The bash command to execute."}}, "required": ["command"]}}}]
deepseek-v4-flash-official
{ "task_id": "task_000006_fd22050f", "domain": "data_querying", "reward": 1, "k_index": 0, "n_bash": 6, "steps_taken": 6, "model": "deepseek-v4-flash" }
[ { "content": "You are a helpful assistant that can interact with a computer.\n\nYour response must include a THOUGHT section before your action where you\nexplain your reasoning. After the THOUGHT, you must call the `bash` tool\nwith EXACTLY ONE bash command (multiple commands chained with `&&` or `||`\ncount a...
[{"type": "function", "function": {"name": "bash", "description": "Execute a bash command. Each command runs in a new subshell.", "parameters": {"type": "object", "properties": {"command": {"type": "string", "description": "The bash command to execute."}}, "required": ["command"]}}}]
deepseek-v4-flash-official
{ "task_id": "task_000009_7a3d643e", "domain": "file_operations", "reward": 1, "k_index": 0, "n_bash": 7, "steps_taken": 7, "model": "deepseek-v4-flash" }
[ { "content": "You are a helpful assistant that can interact with a computer.\n\nYour response must include a THOUGHT section before your action where you\nexplain your reasoning. After the THOUGHT, you must call the `bash` tool\nwith EXACTLY ONE bash command (multiple commands chained with `&&` or `||`\ncount a...
[{"type": "function", "function": {"name": "bash", "description": "Execute a bash command. Each command runs in a new subshell.", "parameters": {"type": "object", "properties": {"command": {"type": "string", "description": "The bash command to execute."}}, "required": ["command"]}}}]
deepseek-v4-flash-official
{ "task_id": "task_000004_3e7c096f", "domain": "scientific_computing", "reward": 1, "k_index": 0, "n_bash": 6, "steps_taken": 6, "model": "deepseek-v4-flash" }
[ { "content": "You are a helpful assistant that can interact with a computer.\n\nYour response must include a THOUGHT section before your action where you\nexplain your reasoning. After the THOUGHT, you must call the `bash` tool\nwith EXACTLY ONE bash command (multiple commands chained with `&&` or `||`\ncount a...
[{"type": "function", "function": {"name": "bash", "description": "Execute a bash command. Each command runs in a new subshell.", "parameters": {"type": "object", "properties": {"command": {"type": "string", "description": "The bash command to execute."}}, "required": ["command"]}}}]
deepseek-v4-flash-official
{ "task_id": "task_000011_a5387288", "domain": "data_science", "reward": 1, "k_index": 0, "n_bash": 6, "steps_taken": 6, "model": "deepseek-v4-flash" }
[ { "content": "You are a helpful assistant that can interact with a computer.\n\nYour response must include a THOUGHT section before your action where you\nexplain your reasoning. After the THOUGHT, you must call the `bash` tool\nwith EXACTLY ONE bash command (multiple commands chained with `&&` or `||`\ncount a...
[{"type": "function", "function": {"name": "bash", "description": "Execute a bash command. Each command runs in a new subshell.", "parameters": {"type": "object", "properties": {"command": {"type": "string", "description": "The bash command to execute."}}, "required": ["command"]}}}]
deepseek-v4-flash-official
{ "task_id": "task_000008_360f148b", "domain": "scientific_computing", "reward": 1, "k_index": 0, "n_bash": 8, "steps_taken": 8, "model": "deepseek-v4-flash" }
[ { "content": "You are a helpful assistant that can interact with a computer.\n\nYour response must include a THOUGHT section before your action where you\nexplain your reasoning. After the THOUGHT, you must call the `bash` tool\nwith EXACTLY ONE bash command (multiple commands chained with `&&` or `||`\ncount a...
[{"type": "function", "function": {"name": "bash", "description": "Execute a bash command. Each command runs in a new subshell.", "parameters": {"type": "object", "properties": {"command": {"type": "string", "description": "The bash command to execute."}}, "required": ["command"]}}}]
deepseek-v4-flash-official
{ "task_id": "task_000008_bebb3084", "domain": "data_science", "reward": 1, "k_index": 0, "n_bash": 7, "steps_taken": 7, "model": "deepseek-v4-flash" }
[{"content":"You are a helpful assistant that can interact with a computer.\n\nYour response must in(...TRUNCATED)
"[{\"type\": \"function\", \"function\": {\"name\": \"bash\", \"description\": \"Execute a bash comm(...TRUNCATED)
deepseek-v4-flash-official
{"task_id":"task_000005_a72a26b5","domain":"file_operations","reward":1,"k_index":0,"n_bash":11,"ste(...TRUNCATED)
End of preview. Expand in Data Studio

TB Lite 终端任务多轮 rollout 轨迹数据集

本数据集为终端任务(TB Lite / terminal-bench 风格)上的 agent 多轮交互轨迹,含 thinking(CoT)字段,用于 SFT 与自蒸馏实验。包含两批数据:DeepSeek 教师采集与 Qwen3.5-9B 自采。两批数据在同一批 2191 个容器环境上采集,采样设置一致。

文件清单

文件 行数 成分 说明
tmax_deepseek_v4flash_multi.parquet 8760 reward=1: 6163 / reward=0: 2597 DeepSeek 教师 k=4 全量,成功与失败轨迹全保留,reward 标注
tmax_deepseek_v4flash_multi_success.parquet 6163 reward=1 上一文件按 reward==1 过滤的子集
tmax_9b_onpolicy_thinking_selfdistill.parquet 4264 reward=1 Qwen3.5-9B 自采 k=4,仅保留通过轨迹

采集设置

项目 DeepSeek 教师 Qwen3.5-9B 自采
模型 deepseek-v4-flash(官方 API) Qwen3.5-9B(vLLM 本地部署)
source 标签 deepseek-v4-flash-official 9b-onpolicy-thinking
temperature / top_p 0.7 / 0.95 0.7 / 0.95
thinking reasoning_effort=high enable_thinking=true(--reasoning-parser qwen3)
k(每环境采样数) 4 4
最大轮数 64 64
环境数 2191 2191
保留策略 成功+失败全留 仅保留通过轨迹

CoT 落在 reasoning_content 字段;工具调用解析使用 qwen3_xml(Qwen 侧)。DeepSeek k=0 轨迹来自更早一次相同参数的单轮采集,k=1–3 为本次续采。

字段结构

messages: list<struct>
  - content:            string
  - reasoning_content:  string   # thinking / CoT
  - role:               string   # system | user | assistant | tool
  - tool_call_ids:      list<string>
  - tool_calls:         list<struct{ function{ arguments{command}, name }, id, type }>
tools:    string                  # JSON 序列化的工具 schema
source:   string                  # 采集来源标签
metadata: struct
  - task_id, domain, model:       string
  - reward:                       int64   # 1=通过, 0=未通过
  - k_index:                      int64   # 采样序号 0..3
  - n_bash, steps_taken:          int64

统计

正确率从 rollout 结果日志计算(含通过/未通过全记录);token / 轮数 / CoT 从训练样本计算,token 用 Qwen3.5-9B tokenizer 统一口径。

正确率

指标 DeepSeek 教师 Qwen3.5-9B 自采
逐 rollout 通过率 70.4% 48.7%
pass@k(2191 任务至少一次通过) 78.2% 64.1%
各 k_index 通过率 69.7%–71.1% 48.0%–49.3%

token 长度(整条样本,含 CoT)

分位 DeepSeek(全量含未通过) Qwen3.5-9B(仅通过)
p50 10618 6492
mean 21471 9029
p95 74056 23002
max 231766 103531

不含 CoT 的 token:DeepSeek p50=5786 / mean=11161;Qwen3.5-9B p50=4484 / mean=6451。

轮数(assistant 发话轮数,上限 64)

分位 DeepSeek Qwen3.5-9B
p50 14 22
mean 22 27
p90 57 55

thinking CoT 长度(reasoning_content token)

指标 DeepSeek Qwen3.5-9B
每样本 CoT · p50 4468 1740
每样本 CoT · mean 10311 2578
每轮 CoT · p50 190 28
每轮 CoT · mean 486 95
含 CoT 的轮占比 98.5% 100.0%

口径说明:DeepSeek 长度分布在全量(含未通过)上计算,Qwen3.5-9B 在仅通过子集上计算;未通过轨迹长度分布更靠上,长度维度的直接对比含此口径差异。

数据脱敏

命令输出中出现的一处采集侧凭证已替换为占位符 <REDACTED_DEEPSEEK_API_KEY>(共 4 条唯一轨迹的工具输出)。基准任务本身包含的示例凭证(如 AWS 文档示例 key、鉴权日志样例、SSH 密钥练习内容)为任务内容,予以保留。

用法

from datasets import load_dataset

ds = load_dataset("Aurorra1123/tmax-selfdistill-datasets", "deepseek_v4flash_multi", split="train")
ds = load_dataset("Aurorra1123/tmax-selfdistill-datasets", "qwen35_9b_onpolicy_thinking", split="train")
Downloads last month
44