Dataset Viewer
Auto-converted to Parquet Duplicate
id
stringlengths
36
36
conversations
listlengths
9
51
tools
stringclasses
6 values
category
stringclasses
8 values
subcategory
stringclasses
23 values
task
stringlengths
51
100
1b510b01-5892-4810-8663-8f457280d904
[ { "from": "system", "value": "You are a function calling AI model. You are provided with function signatures within <tools> </tools> XML tags. You may call one or more functions to assist with the user query. If available tools are not relevant in assisting with user query, just respond in natural conversat...
[{"name": "patch", "description": "Targeted find-and-replace edits in files. Use this instead of sed/awk in terminal. Uses fuzzy matching (9 strategies) so minor whitespace/indentation differences won't break it. Returns a unified diff. Auto-runs syntax checks after editing.\n\nReplace mode (default): find a unique str...
Terminal & Coding
Terminal Tasks
Write a Python script that reads a CSV file, cleans the data by removing duplicates and fixing forma
58be3f1e-49bf-4ec6-846e-041c6f9aac7f
[ { "from": "system", "value": "You are a function calling AI model. You are provided with function signatures within <tools> </tools> XML tags. You may call one or more functions to assist with the user query. If available tools are not relevant in assisting with user query, just respond in natural conversat...
[{"name": "patch", "description": "Targeted find-and-replace edits in files. Use this instead of sed/awk in terminal. Uses fuzzy matching (9 strategies) so minor whitespace/indentation differences won't break it. Returns a unified diff. Auto-runs syntax checks after editing.\n\nReplace mode (default): find a unique str...
Terminal & Coding
Terminal Tasks
Search for Python memory usage techniques, profile the current code, and apply relevant optimization
e8c4ab8d-42c7-4c30-b249-adde1cee4b4c
[ { "from": "system", "value": "You are a function calling AI model. You are provided with function signatures within <tools> </tools> XML tags. You may call one or more functions to assist with the user query. If available tools are not relevant in assisting with user query, just respond in natural conversat...
[{"name": "patch", "description": "Targeted find-and-replace edits in files. Use this instead of sed/awk in terminal. Uses fuzzy matching (9 strategies) so minor whitespace/indentation differences won't break it. Returns a unified diff. Auto-runs syntax checks after editing.\n\nReplace mode (default): find a unique str...
Terminal & Coding
Terminal Tasks
Write a Python script that reads a CSV file, cleans the data by removing duplicates and fixing forma
6641e310-6062-47fc-acc2-89c7d66829f9
[{"from":"system","value":"You are a function calling AI model. You are provided with function signa(...TRUNCATED)
"[{\"name\": \"patch\", \"description\": \"Targeted find-and-replace edits in files. Use this instea(...TRUNCATED)
Terminal & Coding
Terminal Tasks
"Search for Python memory usage techniques, profile the current code, and apply relevant optimizatio(...TRUNCATED)
0c5f862a-7601-4b73-83c1-06c617f20ec6
[{"from":"system","value":"You are a function calling AI model. You are provided with function signa(...TRUNCATED)
"[{\"name\": \"patch\", \"description\": \"Targeted find-and-replace edits in files. Use this instea(...TRUNCATED)
File Operations
File Tasks
Write a Redis caching layer for API responses with configurable TTL.
393d522d-4ff3-44c8-8f06-de798f6bfe22
[{"from":"system","value":"You are a function calling AI model. You are provided with function signa(...TRUNCATED)
"[{\"name\": \"patch\", \"description\": \"Targeted find-and-replace edits in files. Use this instea(...TRUNCATED)
Terminal & Coding
Terminal Tasks
Look up the pytest documentation, install it, configure it for our project, and run it.
8efb11fe-4610-49cb-94ce-96cb6b8c1abd
[{"from":"system","value":"You are a function calling AI model. You are provided with function signa(...TRUNCATED)
"[{\"name\": \"patch\", \"description\": \"Targeted find-and-replace edits in files. Use this instea(...TRUNCATED)
Terminal & Coding
Terminal Tasks
Write unit tests for a Python function that calculates shipping costs. Use pytest.
0a6a4c9a-1993-4dd0-9ec4-d7debea9de50
[{"from":"system","value":"You are a function calling AI model. You are provided with function signa(...TRUNCATED)
"[{\"name\": \"patch\", \"description\": \"Targeted find-and-replace edits in files. Use this instea(...TRUNCATED)
Terminal & Coding
Terminal Tasks
"Write a Rust program that creates a simple chat server and client. Compile/run it and show the outp(...TRUNCATED)
66a5b2ab-d0c2-4bbd-b092-aa9ed1e6995b
[{"from":"system","value":"You are a function calling AI model. You are provided with function signa(...TRUNCATED)
"[{\"name\": \"patch\", \"description\": \"Targeted find-and-replace edits in files. Use this instea(...TRUNCATED)
Terminal & Coding
Terminal Tasks
"Remember that my preferred test framework is pytest with fixtures. Then help me with writing a new (...TRUNCATED)
b6c9ba44-3dd6-4023-8de6-df00decb4d00
[{"from":"system","value":"You are a function calling AI model. You are provided with function signa(...TRUNCATED)
"[{\"name\": \"patch\", \"description\": \"Targeted find-and-replace edits in files. Use this instea(...TRUNCATED)
Terminal & Coding
Terminal Tasks
Use the Docker/Modal skill to set up a cloud environment for large-scale data processing.
End of preview. Expand in Data Studio

lambda/hermes-agent-reasoning-traces(複製)

データの内容

Hermes Agent のハーネスで、実際のツール実行(端末、ファイル操作、ブラウザ)を使って生成した、複数ターンのツール呼び出しの軌跡のデータ。1 行が 1 件の会話で、conversations にステップごとの推論(<think>)と実行結果、category と subcategory にタスクの種類が入る。モデルごとに kimi(Kimi-K2.5)と glm-5.1(GLM-5.1-FP8)の 2 つの config があり、ライセンスは Apache 2.0 である。元モデルの出力の利用規約は未確認である。

データの概要

項目 値
行数 14,701(2 構成・2 分割の合計)
列数 6
容量 1.62 GB(ファイル 6 本)
ファイルの形式 parquet 4
言語 en
公開 public

構成と分割

構成(config) 分割(split)ごとの行数 列数
glm-5.1 train 7,055 6
kimi train 7,646 6

列(構成 glm-5.1・分割 train。どの構成も同じ列)

列 型 先頭の行の値
id string 1b510b01-5892-4810-8663-8f457280d904
conversations list<struct{from: string, value: string}> [29 件]
tools string [{"name": "patch", "description": "Targeted find…
category string Terminal & Coding
subcategory string Terminal Tasks
task string Write a Python script that reads a CSV file, cle…

行数と列は HF のデータセットビューアの集計(元のデータセットの現在の版で数えたもの)。全データセットの一覧は HYUKI5/dataset-catalog にある。

複製について

lambda/hermes-agent-reasoning-traces をリビジョン b92885e4f0161d4b2536512710e004d4892cac6e の時点でそのまま複製したもの。中身は変えていない。判断モデル Laya(convaiinnovations/laya)の微調整に使う候補として、調査メモ(Laya 学習用データセット調査)に載せたデータを集めている。

出典

論文

  • 元のカードに論文へのリンクは見つからなかった

アノテーションの方法

(元のカードに該当の見出しが無いので、ラベルの付け方に触れた段落を抜き出した)

どちらのデータセットも hermes-agent-generator のパイプラインで、実際のツール実行(ターミナルのコマンド、ファイル操作、ブラウザの操作)を使って生成された。合成された出力ではない。

微調整に使うときの注意

  • 調査メモの台帳での区分: B(継承条件・gated・モデル由来の条件・上流の規約などが付く)。区分は記載の文言で機械的に付けたもので、法的な判定ではない。学習に入れる前に元のカードとライセンスの原文で確かめる。
  • Laya の入力は 2,048 token まで(ModernBERT-large / mmBERT)。長い本文は先頭から黙って切られるので、切れる行は事前に除くか、要約・切り出しをしてから入れる。
  • 分割は記事・文書単位で行い、同じ本文から作った別の問いを学習側と較正側に分けない。温度と閾値は学習に使っていない分割で別々に当てる。

調査メモの記述

メモ 節 ライセンス(記載) 確認 注記
agent_trajectories_thought_logs.md Q2. ツール利用・Web・deep research の軌跡で、成否・報酬・品質評価が付いたものは何があるか (L57) apache-2.0 確認済 生成元モデルの出力規約が別に及ぶ(Q4) / 原文注記: 成否・品質ラベルなし。Kimi・GLM の出力利用規約は未確認(事前知識・要確認)

元のカード(日本語訳)

以下は元のデータセットのカードの本文を日本語に訳したもの(機械翻訳)。訳に疑問があるときは、末尾の原文を確かめる。


Hermes Agent の推論トレース

Hermes Agent のハーネスを使って AI エージェントを学習させるための、複数ターンのツール呼び出しの軌跡である。各サンプルは実際のエージェントの会話で、ステップごとの推論(<think> ブロック)と、実際のツール実行結果を含む。

このデータセットには、元になるモデルごとに 2 つの config がある:

Config Model Samples
kimi Moonshot AI Kimi-K2.5 7,646
glm-5.1 ZhipuAI GLM-5.1-FP8 7,055

読み込み

from datasets import load_dataset

# Kimi-K2.5 traces
ds = load_dataset("lambda/hermes-agent-reasoning-traces", "kimi", split="train")

# GLM-5.1 traces
ds = load_dataset("lambda/hermes-agent-reasoning-traces", "glm-5.1", split="train")

スキーマ

2 つの config は同じスキーマを共有する:

Field Type Description
id string UUID による識別子
conversations list 複数ターンの対話(system、human、gpt、tool のメッセージ)
tools string エージェントが使えるツールの定義(JSON)
category string 上位のタスクのカテゴリ
subcategory string より細かいタスクの種類
task string タスクの説明(ユーザーのプロンプトから)

会話のメッセージは ShareGPT 形式を使う:

{"from": "system|human|gpt|tool", "value": "..."}
  • <think> ブロックには思考連鎖による推論が入る
  • <tool_call> ブロックには関数の呼び出しが入る
  • <tool_response> ブロックには実際の実行結果が入る

統計

Metric kimi glm-5.1
サンプル数 7,646 7,055
総ターン数 185,798 134,918
ツール呼び出しの総数 106,222 68,328
サンプルあたりの平均ターン数 24.3 19.1
サンプルあたりの平均ツール呼び出し数 13.9 9.7
<think> の平均の深さ(語数) 414 70

カテゴリ

2 つの config は、共通の 9 カテゴリの分類を使う:

Category kimi glm-5.1
ターミナルとコーディング 2,010 2,237
エージェントのツール 1,474 2,775
リポジトリのタスク 1,109 1,022
ブラウザの自動化 1,048 639
マルチツール 807 52
ファイル操作 757 134
スケジューリング 204 104
計画と整理 201 92
会話 36 0

生成の詳細

Kimi-K2.5

  • モデル: moonshotai/Kimi-K2.5(MoE)
  • 推論: vLLM を --tool-call-parser kimi_k2 --reasoning-parser kimi_k2 --enable-auto-tool-choice を付けて使用

GLM-5.1

  • モデル: zai-org/GLM-5.1-FP8
  • 推論: vLLM を --tool-call-parser glm47 --reasoning-parser glm45 --enable-auto-tool-choice を付けて使用
  • 提供: 負荷分散ゲートウェイ経由の 8xH100 ノード 3 台(3x 8xH100)
  • コンテキスト: 最大 202,752 token。MTP の投機的デコード

どちらのデータセットも hermes-agent-generator のパイプラインで、実際のツール実行(ターミナルのコマンド、ファイル操作、ブラウザの操作)を使って生成された。合成された出力ではない。

データの出典

どちらのデータセットも、同じタスクのカテゴリにまたがる軌跡を含む:

  • ターミナルとコーディング — スクリプトの作成、デバッグ、環境構築、データ処理、テスト、ドキュメント作成
  • ブラウザの自動化 — Playwright による移動、スクレイピング、フォーム入力、スクリーンショットの分析
  • エージェントのツール — Hermes 固有の機能: メモリの永続化、タスクの委譲、スキルの管理、ToDo の計画、コードの実行、セッションの想起
  • リポジトリのタスク — GitHub のリポジトリにまたがる実際のコードベースでの作業: バグ修正、機能の実装、テストの作成、コードレビュー、リファクタリング

ライセンス

Apache 2.0

原文(英語)

Hermes Agent Reasoning Traces

Multi-turn tool-calling trajectories for training AI agents using the Hermes Agent harness. Each sample is a real agent conversation with step-by-step reasoning (<think> blocks) and actual tool execution results.

This dataset has two configs, one per source model:

Config Model Samples
kimi Moonshot AI Kimi-K2.5 7,646
glm-5.1 ZhipuAI GLM-5.1-FP8 7,055

Loading

from datasets import load_dataset

# Kimi-K2.5 traces
ds = load_dataset("lambda/hermes-agent-reasoning-traces", "kimi", split="train")

# GLM-5.1 traces
ds = load_dataset("lambda/hermes-agent-reasoning-traces", "glm-5.1", split="train")

Schema

Both configs share the same schema:

Field Type Description
id string UUID identifier
conversations list Multi-turn dialogue (system, human, gpt, tool messages)
tools string JSON tool definitions available to the agent
category string High-level task category
subcategory string Fine-grained task type
task string Task description (from user prompt)

Conversation messages use ShareGPT format:

{"from": "system|human|gpt|tool", "value": "..."}
  • <think> blocks contain chain-of-thought reasoning
  • <tool_call> blocks contain function invocations
  • <tool_response> blocks contain real execution results

Statistics

Metric kimi glm-5.1
Samples 7,646 7,055
Total turns 185,798 134,918
Total tool calls 106,222 68,328
Avg turns per sample 24.3 19.1
Avg tool calls per sample 13.9 9.7
Avg <think> depth (words) 414 70

Categories

Both configs use a shared 9-category taxonomy:

Category kimi glm-5.1
Terminal & Coding 2,010 2,237
Agent Tools 1,474 2,775
Repository Tasks 1,109 1,022
Browser Automation 1,048 639
Multi-Tool 807 52
File Operations 757 134
Scheduling 204 104
Planning & Organization 201 92
Conversational 36 0

Generation Details

Kimi-K2.5

  • Model: moonshotai/Kimi-K2.5 (MoE)
  • Inference: vLLM with --tool-call-parser kimi_k2 --reasoning-parser kimi_k2 --enable-auto-tool-choice

GLM-5.1

  • Model: zai-org/GLM-5.1-FP8
  • Inference: vLLM with --tool-call-parser glm47 --reasoning-parser glm45 --enable-auto-tool-choice
  • Serving: 3x 8xH100 nodes via load-balanced gateway
  • Context: 202,752 tokens max, MTP speculative decoding

Both datasets were generated using the hermes-agent-generator pipeline with real tool execution (terminal commands, file operations, browser actions) — not synthetic outputs.

Data Sources

Both datasets include trajectories across the same task categories:

  • Terminal & Coding — script writing, debugging, environment setup, data processing, testing, documentation
  • Browser Automation — Playwright-based navigation, scraping, form filling, screenshot analysis
  • Agent Tools — Hermes-specific capabilities: memory persistence, task delegation, skill management, todo planning, code execution, session recall
  • Repository Tasks — real codebase work across GitHub repos: bug fixes, feature implementation, test writing, code review, refactoring

License

Apache 2.0

Downloads last month
60