Dataset Preview
Duplicate
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed
Error code:   DatasetGenerationError
Exception:    ValueError
Message:      Unmatched ''"' when when decoding 'string'
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1835, in _prepare_split_single
                  for key, table in generator:
                                    ^^^^^^^^^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 621, in wrapped
                  for item in generator(*args, **kwargs):
                              ~~~~~~~~~^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 200, in _generate_tables
                  parse_traces_info(lines)
                  ~~~~~~~~~~~~~~~~~^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 448, in parse_traces_info
                  decoded_event = ujson_loads(event)
                File "/usr/local/lib/python3.14/site-packages/datasets/utils/json.py", line 20, in ujson_loads
                  return pd.io.json.ujson_loads(*args, **kwargs)
                         ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
              ValueError: Unmatched ''"' when when decoding 'string'
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1379, in compute_config_parquet_and_info_response
                  parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
                                                                        ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      builder, max_dataset_size_bytes=max_dataset_size_bytes
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 956, in stream_convert_to_parquet
                  builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
                  ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1702, in _prepare_split
                  for job_id, done, content in self._prepare_split_single(
                                               ~~~~~~~~~~~~~~~~~~~~~~~~~~^
                      gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  ):
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1887, in _prepare_split_single
                  raise DatasetGenerationError("An error occurred while generating the dataset") from e
              datasets.exceptions.DatasetGenerationError: An error occurred while generating the dataset

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

harness
string
session_id
string
prompt
string
messages
list
tools
list
metadata
unknown
sent_at
string
num_user_messages
int64
num_tool_calls
int64
trace
unknown
file_path
string
pi
cfa672f2-6e86-43d2-ba2e-d31354d8418f
use github cli and checkout pull request 736
[ { "role": "user", "content": "use github cli and checkout pull request 736" }, { "role": "assistant", "content": "", "tool_calls": [ { "id": "toolu_01LNCFiEFT8QSXL3Jz4n9pQZ", "type": "function", "function": { "name": "bash", "arguments": { ...
[ { "type": "function", "function": { "name": "bash", "parameters": { "type": "object", "properties": { "command": { "type": "string" } }, "additionalProperties": true, "required": [ "command" ] } }...
{ "source_file": "LarsEckart-approvaltests-java-sessions-2025-12-01T17-10-20-593Z_cfa672f2-6e86-43d2-ba2e-d31354d8418f.jsonl", "session_id": "cfa672f2-6e86-43d2-ba2e-d31354d8418f", "trace_type": "pi", "model_provider": null, "model": null, "cwd": "[PROJECT_ROOT]", "cli_version": null, "turn_count": 2, ...
2025-12-01T17:10:35.051Z
2
3
"{\"type\":\"session\",\"id\":\"cfa672f2-6e86-43d2-ba2e-d31354d8418f\",\"timestamp\":\"2025-12-01T17:10:38.702Z\",\"cwd\":\"[PROJECT_ROOT]\",\"provider\":\"anthropic\",\"modelId\":\"claude-opus-4-5\",\"thinkingLevel\":\"off\"}\n{\"type\":\"message\",\"timestamp\":\"2025-12-01T17:10:35.051Z\",\"message\":{\"role\":\"use...
LarsEckart-approvaltests-java-sessions-2025-12-01T17-10-20-593Z_cfa672f2-6e86-43d2-ba2e-d31354d8418f.jsonl
pi
b698a33a-749d-4f5f-abbe-4360a59354b6
use github cli, read issue 743 and then Thoroughly help me find the root cause of the issue and don't stop investigating until you're confident you've found one or multiple reasons for this issue.
[{"role":"user","content":"use github cli, read issue 743 and then Thoroughly help me find the root (...TRUNCATED)
[{"type":"function","function":{"name":"bash","parameters":{"type":"object","properties":{"command":(...TRUNCATED)
{"source_file":"LarsEckart-approvaltests-java-sessions-2025-12-04T08-54-31-360Z_b698a33a-749d-4f5f-a(...TRUNCATED)
2025-12-04T08:54:49.162Z
2
49
"{\"type\":\"session\",\"id\":\"b698a33a-749d-4f5f-abbe-4360a59354b6\",\"timestamp\":\"2025-12-04T08(...TRUNCATED)
LarsEckart-approvaltests-java-sessions-2025-12-04T08-54-31-360Z_b698a33a-749d-4f5f-abbe-4360a59354b6.jsonl
pi
4fb30a4c-bcf9-447e-9ae4-97fe6c91a666
use github cli and read issue 743 . describe a test scenario that would highlight all the issues shown there. enumerate the issues shown.
[{"role":"user","content":"use github cli and read issue 743 . describe a test scenario that would h(...TRUNCATED)
[{"type":"function","function":{"name":"bash","parameters":{"type":"object","properties":{"command":(...TRUNCATED)
{"source_file":"LarsEckart-approvaltests-java-sessions-2025-12-08T17-18-45-938Z_4fb30a4c-bcf9-447e-9(...TRUNCATED)
2025-12-08T17:19:29.243Z
1
1
"{\"type\":\"session\",\"id\":\"4fb30a4c-bcf9-447e-9ae4-97fe6c91a666\",\"timestamp\":\"2025-12-08T17(...TRUNCATED)
LarsEckart-approvaltests-java-sessions-2025-12-08T17-18-45-938Z_4fb30a4c-bcf9-447e-9ae4-97fe6c91a666.jsonl
pi
53dd8ca6-78bf-4d55-b3a5-88515fe4cb20
rename all files in @internal_documentation/micro_features/ to use camelCase
[{"role":"user","content":"rename all files in @internal_documentation/micro_features/ to use camelC(...TRUNCATED)
[{"type":"function","function":{"name":"bash","parameters":{"type":"object","properties":{"command":(...TRUNCATED)
{"source_file":"LarsEckart-approvaltests-java-sessions-2025-12-15T17-39-23-937Z_53dd8ca6-78bf-4d55-b(...TRUNCATED)
2025-12-15T17:41:19.879Z
1
2
"{\"type\":\"session\",\"id\":\"53dd8ca6-78bf-4d55-b3a5-88515fe4cb20\",\"timestamp\":\"2025-12-15T17(...TRUNCATED)
LarsEckart-approvaltests-java-sessions-2025-12-15T17-39-23-937Z_53dd8ca6-78bf-4d55-b3a5-88515fe4cb20.jsonl
pi
fa32e25f-0321-46de-9ff7-5a9c8f07ea89
read @internal_documentation/micro_features/firstWorkingReporterToString.md and implement it.
[{"role":"user","content":"read @internal_documentation/micro_features/firstWorkingReporterToString.(...TRUNCATED)
[{"type":"function","function":{"name":"bash","parameters":{"type":"object","properties":{"command":(...TRUNCATED)
{"source_file":"LarsEckart-approvaltests-java-sessions-2025-12-15T17-42-28-899Z_fa32e25f-0321-46de-9(...TRUNCATED)
2025-12-15T17:42:44.192Z
4
40
"{\"type\":\"session\",\"id\":\"fa32e25f-0321-46de-9ff7-5a9c8f07ea89\",\"timestamp\":\"2025-12-15T17(...TRUNCATED)
LarsEckart-approvaltests-java-sessions-2025-12-15T17-42-28-899Z_fa32e25f-0321-46de-9ff7-5a9c8f07ea89.jsonl
pi
42d2afba-2418-4626-9f22-78b11e682608
read @internal_documentation/micro_features/DateScrubber.md and implement it.
[{"role":"user","content":"read @internal_documentation/micro_features/DateScrubber.md and implement(...TRUNCATED)
[{"type":"function","function":{"name":"bash","parameters":{"type":"object","properties":{"command":(...TRUNCATED)
{"source_file":"LarsEckart-approvaltests-java-sessions-2025-12-15T17-59-06-084Z_42d2afba-2418-4626-9(...TRUNCATED)
2025-12-15T17:59:14.235Z
1
19
"{\"type\":\"session\",\"id\":\"42d2afba-2418-4626-9f22-78b11e682608\",\"timestamp\":\"2025-12-15T17(...TRUNCATED)
LarsEckart-approvaltests-java-sessions-2025-12-15T17-59-06-084Z_42d2afba-2418-4626-9f22-78b11e682608.jsonl
pi
0ac12b5e-c363-49f5-976e-885df81a2087
read @approvaltests-util/src/main/java/com/spun/util/io/FileUtils.java method writeFile and tell me why this might fail. we're having a failing test and it doesnt create the recieved file
[{"role":"user","content":"read @approvaltests-util/src/main/java/com/spun/util/io/FileUtils.java me(...TRUNCATED)
[{"type":"function","function":{"name":"bash","parameters":{"type":"object","properties":{"command":(...TRUNCATED)
{"source_file":"LarsEckart-approvaltests-java-sessions-2025-12-15T18-14-58-124Z_0ac12b5e-c363-49f5-9(...TRUNCATED)
2025-12-15T18:30:06.983Z
3
39
"{\"type\":\"session\",\"id\":\"0ac12b5e-c363-49f5-976e-885df81a2087\",\"timestamp\":\"2025-12-15T18(...TRUNCATED)
LarsEckart-approvaltests-java-sessions-2025-12-15T18-14-58-124Z_0ac12b5e-c363-49f5-976e-885df81a2087.jsonl
pi
854baca6-1dd3-4e35-b9ee-8970241f3b38
we made PackageSettings with a private constructor (see git status which test) and now in @approvaltests-util/src/main/java/org/packagesettings/PackageLevelSettings.java throws an exception: "java.lang.IllegalAccessException: class org.packagesettings.PackageLevelSettings cannot access a member of class org.approvaltes...
[{"role":"user","content":"we made PackageSettings with a private constructor (see git status which (...TRUNCATED)
[{"type":"function","function":{"name":"bash","parameters":{"type":"object","properties":{"command":(...TRUNCATED)
{"source_file":"LarsEckart-approvaltests-java-sessions-2026-01-05T17-35-20-945Z_854baca6-1dd3-4e35-b(...TRUNCATED)
2026-01-05T17:35:48.851Z
15
156
"{\"type\":\"session\",\"version\":3,\"id\":\"854baca6-1dd3-4e35-b9ee-8970241f3b38\",\"timestamp\":\(...TRUNCATED)
LarsEckart-approvaltests-java-sessions-2026-01-05T17-35-20-945Z_854baca6-1dd3-4e35-b9ee-8970241f3b38.jsonl
pi
628c5933-e5b8-4bbd-8b1b-8bb4513c57d7
hey, i just merged PR #748 we have to pull those changes and then they had a package name wrong, they used `_`, let's fix that
[{"role":"user","content":"hey, i just merged PR #748 we have to pull those changes and then they ha(...TRUNCATED)
[{"type":"function","function":{"name":"bash","parameters":{"type":"object","properties":{"command":(...TRUNCATED)
{"source_file":"LarsEckart-approvaltests-java-sessions-2026-01-12T17-25-23-276Z_628c5933-e5b8-4bbd-8(...TRUNCATED)
2026-01-12T17:25:43.390Z
20
151
"{\"type\":\"session\",\"version\":3,\"id\":\"628c5933-e5b8-4bbd-8b1b-8bb4513c57d7\",\"timestamp\":\(...TRUNCATED)
LarsEckart-approvaltests-java-sessions-2026-01-12T17-25-23-276Z_628c5933-e5b8-4bbd-8b1b-8bb4513c57d7.jsonl
pi
063ee7fa-f8e1-4ab2-bccc-fdd452dbcf9e
use github cli and read https://github.com/approvals/ApprovalTests.Java/issues/112 there are comments which I haven't addressed yet
[{"role":"user","content":"use github cli and read https://github.com/approvals/ApprovalTests.Java/i(...TRUNCATED)
[{"type":"function","function":{"name":"bash","parameters":{"type":"object","properties":{"command":(...TRUNCATED)
{"source_file":"LarsEckart-approvaltests-java-sessions-2026-01-19T18-13-01-383Z_063ee7fa-f8e1-4ab2-b(...TRUNCATED)
2026-01-19T18:13:16.324Z
3
25
"{\"type\":\"session\",\"version\":3,\"id\":\"063ee7fa-f8e1-4ab2-bccc-fdd452dbcf9e\",\"timestamp\":\(...TRUNCATED)
LarsEckart-approvaltests-java-sessions-2026-01-19T18-13-01-383Z_063ee7fa-f8e1-4ab2-bccc-fdd452dbcf9e.jsonl
End of preview.

MaxDevv/real-pi-coding-agent-traces-sessions(複製)

データの内容

実在の人がコーディングエージェント pi を使い、実在のオープンソースソフトウェアを作った際の実際のセッションを集めたデータ。1 つの .jsonl ファイルが 1 セッションで、ユーザーのプロンプト・アシスタントの応答・ツール呼び出し(read・bash・edit・write など)・ツールの結果・思考ブロックが時系列に入っている。21 の公開データセットから集め、合成データや AI が生成した内容は人手で除いたとカードにある。元データごとの権利は未確認である。

データの概要

項目 値
行数・列数 不明(データセットビューアで読めない形式。ファイルの内訳だけを示す)
容量 777.89 MB(ファイル 1294 本)
ファイルの形式 jsonl 198
公開 public

全データセットの一覧は HYUKI5/dataset-catalog にある。

複製について

MaxDevv/real-pi-coding-agent-traces-sessions をリビジョン 8c593252ddad7dca08a0afc07896195fa73f2d6e 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)

21 件の独立に公開された Hugging Face のデータセットから集め、合成データや AI が生成したコンテンツを人手で除外した、実際の人間と AI によるコーディングエージェントのセッションの集約データセットである。

Hugging Face 上の「エージェントの対話」データセットの多くは合成データである。LLM が生成したプロンプト、模擬ユーザー、テンプレート化されたタスク、台本化されたシナリオから成る。これらはいくつかの用途には有用だが、実際のエージェントのトレースを価値あるものにする性質を体系的に欠いている:

  1. 発見 — pi-share-hf で書き出されたすべてのデータセットを、Hugging Face で全文検索した。
  2. 人手による検査 — 候補のデータセットをすべて開いて読んだ。AI が生成したプロンプト、模擬ユーザー、テンプレート化された合成タスクからなるデータセットは除外した。実在のリポジトリに対する、実際の人間とエージェントの対話のデータセットのみを残した。合成プロンプトや模擬ユーザーの振る舞いを含む候補は、この段階で不採用とした。
  3. 重複の除去 — 内容のハッシュ(ファイルのバイト列の MD5)によりセッションの重複を除いた。バイト単位で同一のものを除いた。全体で重複は 1 件除かれた。
  4. 集約 — 1 つのデータセットにまとめた。セッションごとの出所は manifest.jsonl に残している。

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 Q5. 実利用の会話ログと、思考ログを含む公開物には、どんなラベルが付いているか (L152) MaxDevv は other(HF 表記)。trace-commons は cc-by-4.0 確認済 生成元モデルの出力規約が別に及ぶ(Q4) / 原文注記: 元の 21 データの個別の権利は未確認で、集約物の other は条文を取得していない。1,291 件と小さい。ユーザーのプロンプトに個人情報が残る可能性(削除済みとカードは述べるが検証していない)

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。


実際の Pi コーディングエージェントのトレース・セッション

21 件の独立に公開された Hugging Face のデータセットから集め、合成データや AI が生成したコンテンツを人手で除外した、実際の人間と AI によるコーディングエージェントのセッションの集約データセットである。

各セッションは、実在の人物が pi(オープンソースの AI コーディングエージェント基盤)を使って、実在のオープンソースソフトウェアを構築・デバッグ・リリースした際の、編集されていない(ただし墨消し済みの)トレースである。実際のプロンプト、実際のツール呼び出し、実際のエラー、実際の後戻りが含まれる。

このデータセットが存在する理由

Hugging Face 上の「エージェントの対話」データセットの多くは合成データである。LLM が生成したプロンプト、模擬ユーザー、テンプレート化されたタスク、台本化されたシナリオから成る。これらはいくつかの用途には有用だが、実際のエージェントのトレースを価値あるものにする性質を体系的に欠いている:

  • 実際の問題設定 — 曖昧で誤字を含み、文脈に依存する人間のプロンプトであり、きれいに仕様化されたタスクではない
  • 実際の失敗と回復 — ツールのエラー、誤った前提、行き止まり、後戻り、うまくいかなかった設計の反復
  • 実際の複雑さ — 複数ファイルにまたがるリファクタリング、API の癖、環境の問題、型エラー、dependency hell
  • 実際のツール利用パターン — 実在のコードベースに対する read / bash / edit / write 呼び出しの実際の系列と、実際の出力
  • 実際の推論 — 推論モデルが本物の曖昧さを検討した際の思考ブロック

コーディングエージェントのための蒸留、SFT、軌跡レベルの学習を行うなら、合成データと実データの差は重要である。このデータセットは、多数の小さな単一プロジェクトのデータセットに散らばって既に存在する実データを、一箇所にまとめる試みである。

データセットの概要

指標 値
セッション総数 1,291
元データセット数 21
メッセージ・ターンの総数 192,309
おおよそのトークン数 ~194,000,000
セッションの日付の範囲 2025-12-01 → 2026-06-18
形式 改行区切りの JSON(pi のセッション形式)
主な言語 プロンプトは英語。コードは TypeScript、Java、Python、Lua など

セッションの中身

各 *.jsonl ファイルは、pi のセッション形式で書かれた 1 つの完全なエージェントのセッションである。改行区切りの JSON で、次を含む:

内容 説明
ユーザーのプロンプト 人間がエージェントに実際に頼んだこと
アシスタントの応答 コード、説明、分析、計画
ツール呼び出し read、bash、edit、write、およびカスタムの拡張ツール。実際に行われた操作
ツールの結果 コマンドの出力、ファイルの内容、コンパイラのエラー、テストの失敗
思考ブロック モデルの推論(推論モデルを使った場合のみ)

模式的な例(例示のため省略している):

{"type": "message", "role": "user", "content": "the websocket gateway drops connections after ~30s, figure out why"}
{"type": "message", "role": "assistant", "content": [{"type": "thinking", "thinking": "..."}, {"type": "text", "text": "Let me look at the gateway code..."}, {"type": "tool_call", "name": "read", "arguments": {"path": "src/gateway.ts"}}]}
{"type": "tool_result", "name": "read", "content": "..."}

セッションは、実際の開発作業の広い範囲にわたる。アーキテクチャ設計、機能の実装、デバッグ、テストの修正、連携作業(Discord ボット、WebSocket ゲートウェイ、Ollama プロバイダ)、ツール開発、そして結論を定めない試行錯誤まで含まれる。

読み込み方法

from datasets import load_dataset

ds = load_dataset("MaxDevv/real-pi-coding-agent-traces-sessions")

または個々のセッションファイルを直接取得してもよい。各ファイルは自己完結した NDJSON であり、任意の JSON ライブラリで 1 行ずつ解析できる。

出典と帰属

このデータセットは、以下の公開データセットを集めたものである。元のトレースの功績はすべて、元の作者に帰属する。セッションごとの出所(元データセット、元のファイル名、SHA-256、ライセンス)は manifest.jsonl に記録されている。

選別と処理の方法

  1. 発見 — pi-share-hf で書き出されたすべてのデータセットを、Hugging Face で全文検索した。
  2. 人手による検査 — 候補のデータセットをすべて開いて読んだ。AI が生成したプロンプト、模擬ユーザー、テンプレート化された合成タスクからなるデータセットは除外した。実在のリポジトリに対する、実際の人間とエージェントの対話のデータセットのみを残した。合成プロンプトや模擬ユーザーの振る舞いを含む候補は、この段階で不採用とした。
  3. 重複の除去 — 内容のハッシュ(ファイルのバイト列の MD5)によりセッションの重複を除いた。バイト単位で同一のものを除いた。全体で重複は 1 件除かれた。
  4. 集約 — 1 つのデータセットにまとめた。セッションごとの出所は manifest.jsonl に残している。

元の作者が行った以上の、セッション内容へのさらなるフィルタリング、編集、「クリーニング」は行っていない。雑然さこそが要点である。

マニフェスト

manifest.jsonl には、セッションごとに 1 行が含まれる:

{
  "file": "badlogicgames-pi-mono-2026-01-24T22-14-19-387Z_da2953ac-848d-46fa-acfb-1203d2623538.jsonl",
  "source_dataset": "badlogicgames/pi-mono",
  "source_file": "badlogicgames-pi-mono-2026-01-24T22-14-19-387Z_da2953ac-848d-46fa-acfb-1203d2623538.jsonl",
  "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
  "license": "other",
  "size_bytes": 3002
}

プライバシーと安全性

元の各セッションは pi-share-hf を通じて公開されたものである。pi-share-hf は、何かがアップロードされる前に三層のパイプラインを適用する:

  1. 決定的な墨消し — env ファイル、.npmrc、シェルの設定、利用者が指定した秘密情報のリストに含まれる秘密の値そのものを、[REDACTED] に置き換える
  2. TruffleHog によるスキャン — 墨消し後の出力を、残っている秘密情報がないかスキャンする。検出されたもの(検証済み・未検証・不明のいずれも)は、そのセッションを自動的にブロックする
  3. LLM によるレビュー — 各セッションについて、非公開のインフラ、個人情報、プロジェクト外の内容がないかを確認する。プロジェクトの文脈を用いて、OSS の作業とそれ以外を区別する

この集約によって新たな生データが加わることはない。ここにあるものはすべて、元の段階で既にこれらのチェックを通っている。とはいえ、墨消しのパイプラインが完璧なものはない。いずれかのセッションに秘密情報、認証情報、個人情報を見つけた場合は、直ちに Issue を立てるか連絡すること。該当箇所は削除される(元の公開元にも報告される)。

ライセンス

これは集めたものであり、各セッションはそれぞれの元データセットのライセンスに従う。ライセンスの本文は元データセットのリポジトリを参照すること。元の作者で、自分のセッションを削除したい、または帰属表記を変えたい場合は、Issue を立てれば速やかに対応する。

想定される用途

  • 実際の軌跡からのコーディングエージェントの SFT・蒸留
  • 人間と AI の協働、エージェントの振る舞い、ツール利用パターンに関する研究
  • 現実的な失敗と回復の動態を用いたエージェントの学習と評価
  • 人がコーディングエージェントに実際にどのようにプロンプトを与え、導くかの分析
  • 実際のタスクに基づくベンチマークの構築

対象外の用途

  • セッションの作者を特定し、匿名性を解除し、またはプロファイルすること
  • 残存する秘密情報や個人情報(PII)を抽出すること(「プライバシーと安全性」の項を参照。見つけた場合は報告すること)
  • これらのトレースを自分自身の成果として示すこと

制約と偏り

  • エコシステムの偏り — すべてが pi のセッションである。ツールのスキーマ、プロンプトの書き方、ワークフローは、その基盤に特有のものを反映している
  • 作者の偏り — 少数の開発者がセッションの大半を占める。これは「プログラマー」の代表的な標本ではない
  • 言語・技術スタックの偏り — TypeScript/Bun に大きく偏っている。プロンプトは英語
  • モデルの偏り — アシスタントの振る舞いは、使われた特定のモデルを反映している(kimi-k2.5、glm-5.1、qwen3-coder、gpt-5.x など。セッションと日付によって異なる)
  • 選択の偏り — ここにあるのは作者が公開すると決めたセッションだけである。人は、最も恥ずかしい、または最も機微なデバッグのセッションは公開しない
  • 小規模 — ニッチで忠実度の高いデータセットであり、Web 規模のコーパスではない

引用

@dataset{pi_agent_sessions_traces,
  title   = {Real Pi Coding Agent Traces Sessions},
  author  = {Maximus Adeola},
  year    = {2026},
  url     = {https://huggingface.co/datasets/MaxDevv/real-pi-coding-agent-traces-sessions},
  note    = {Compilation of 21 pi-share-hf session datasets; see manifest.jsonl for per-session attribution}
}

元の作者(上の出典の表を参照)にも、あわせてクレジットを記すこと。作業を実際に行い、トレースを公開することを選んだのは元の作者である。

謝辞と免責事項

pi-share-hf を通じてセッションを公開しているすべての人に感謝する。また、このような種類のデータセットを可能にしている pi と pi-share-hf のパイプラインを作った Mario Zechner (badlogic) に感謝する。

これは非公式のコミュニティによる集約である。pi、その作者、元データセットのいずれの作者とも提携しておらず、支持や後援も受けていない。すべての商標およびプロジェクト名は、それぞれの所有者に帰属する。

原文(英語)

Real Pi Coding Agent Traces Sessions

An aggregated dataset of real human–AI coding agent sessions, collected from 21 independently published Hugging Face datasets and hand-filtered to exclude synthetic or AI-generated content.

Every session is an unedited (but redacted) trace of a real person using pi — an open-source AI coding agent harness — to build, debug, and ship real open-source software. Real prompts, real tool calls, real errors, real backtracking.

Why this dataset exists

Most "agent interaction" datasets on Hugging Face are synthetic: LLM-generated prompts, simulated users, templated tasks, scripted scenarios. They're useful for some things, but they systematically lack the properties that make real agent traces valuable:

  • Real problem statements — vague, typo-ridden, context-dependent human prompts, not cleanly specified tasks
  • Real failure and recovery — tool errors, wrong assumptions, dead ends, backtracking, iterating on a design that didn't work
  • Real complexity — multi-file refactors, API quirks, environment issues, type errors, dependency hell
  • Real tool-use patterns — actual sequences of read / bash / edit / write calls against real codebases, with real outputs
  • Real reasoning — thinking blocks from reasoning models working through genuine ambiguity

If you're doing distillation, SFT, or trajectory-level training for coding agents, the gap between synthetic and real data matters. This dataset is an attempt to gather the real supply that already exists, scattered across many small single-project datasets, into one place.

Dataset at a glance

Stat Value
Total sessions 1,291
Source datasets 21
Total messages / turns 192,309
Approx. tokens ~194,000,000
Date range of sessions 2025-12-01 → 2026-06-18
Format newline-delimited JSON (pi session format)
Primary languages English prompts; TypeScript, Java, Python, Lua, and more in code

What's in a session

Each *.jsonl file is one complete agent session in pi session format: newline-delimited JSON containing:

Content Description
User prompts What the human actually asked the agent to do
Assistant responses Code, explanations, analysis, plans
Tool calls read, bash, edit, write, and custom extension tools — the actual operations performed
Tool results Command output, file contents, compiler errors, test failures
Thinking blocks Model reasoning (where reasoning models were used)

Schematic (illustrative, abbreviated):

{"type": "message", "role": "user", "content": "the websocket gateway drops connections after ~30s, figure out why"}
{"type": "message", "role": "assistant", "content": [{"type": "thinking", "thinking": "..."}, {"type": "text", "text": "Let me look at the gateway code..."}, {"type": "tool_call", "name": "read", "arguments": {"path": "src/gateway.ts"}}]}
{"type": "tool_result", "name": "read", "content": "..."}

Sessions span a wide range of real development work: architecture design, feature implementation, debugging, test fixing, integration work (Discord bots, WebSocket gateways, Ollama providers), tooling development, and open-ended experimentation.

Loading

from datasets import load_dataset

ds = load_dataset("MaxDevv/real-pi-coding-agent-traces-sessions")

Or grab individual session files directly — each is self-contained NDJSON and can be parsed line by line with any JSON library.

Sources & attribution

This dataset is a compilation of the following public datasets. All credit for the underlying traces goes to the original authors. Per-session provenance (source dataset, original filename, SHA-256, license) is recorded in manifest.jsonl.

Curation & processing methodology

  1. Discovery — full-text search of Hugging Face for all datasets exported with pi-share-hf.
  2. Manual inspection — every candidate dataset was opened and read. Datasets consisting of AI-generated prompts, simulated users, or templated/synthetic tasks were excluded. Only datasets of genuine human–agent interaction on real repositories were kept. Candidates were rejected at this stage if they contained synthetic prompts or simulated user behavior.
  3. Deduplication — sessions deduplicated by content hash (MD5 of file bytes), removing byte-identical duplicates. 1 duplicate was removed across the full set.
  4. Aggregation — merged into a single dataset with per-session provenance retained in manifest.jsonl.

No additional filtering, editing, or "cleaning" of session content was applied beyond what upstream authors did. The messiness is the point.

Manifest

manifest.jsonl contains one row per session:

{
  "file": "badlogicgames-pi-mono-2026-01-24T22-14-19-387Z_da2953ac-848d-46fa-acfb-1203d2623538.jsonl",
  "source_dataset": "badlogicgames/pi-mono",
  "source_file": "badlogicgames-pi-mono-2026-01-24T22-14-19-387Z_da2953ac-848d-46fa-acfb-1203d2623538.jsonl",
  "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
  "license": "other",
  "size_bytes": 3002
}

Privacy & safety

Every upstream session was published through pi-share-hf, which applies a three-layer pipeline before anything is uploaded:

  1. Deterministic redaction — exact secret values from env files, .npmrc, shell configs, and user-supplied secret lists are replaced with [REDACTED]
  2. TruffleHog scan — the redacted output is scanned for surviving secrets; any finding (verified, unverified, or unknown) automatically blocks the session
  3. LLM review — checks each session for private infrastructure, personal information, and non-project content, with project context to distinguish OSS work from anything else

This aggregation adds no new raw data — everything here already passed those checks upstream. That said, no redaction pipeline is perfect. If you find a secret, credential, or personal information in any session, please open an issue or contact me immediately and it will be removed (and reported upstream).

Licensing

This is a compilation; each session remains under the license of its source dataset. Consult the source dataset repos for license texts. If you are a source author and want your sessions removed or reattributed, open an issue and it will be handled promptly.

Intended uses

  • SFT / distillation of coding agents from real trajectories
  • Research on human–AI collaboration, agent behavior, tool-use patterns
  • Training and evaluating agents on realistic failure-and-recovery dynamics
  • Analysis of how people actually prompt and steer coding agents
  • Building benchmarks grounded in real tasks

Out-of-scope uses

  • Attempting to identify, de-anonymize, or profile session authors
  • Extracting residual secrets or PII (see Privacy — report instead)
  • Representing these traces as your own work

Limitations & biases

  • Ecosystem bias — everything is pi sessions; tool schemas, prompting styles, and workflows reflect that harness specifically
  • Author bias — a small number of developers account for most sessions; this is not a representative sample of "programmers"
  • Language/stack bias — heavy TypeScript/Bun skew; English prompts
  • Model bias — assistant behavior reflects the specific models used (kimi-k2.5, glm-5.1, qwen3-coder, gpt-5.x, and others, varying by session and date)
  • Selection bias — only sessions whose authors chose to publish are here; people don't publish their most embarrassing or most sensitive debugging sessions
  • Small scale — this is a niche, high-fidelity dataset, not a web-scale corpus

Citation

@dataset{pi_agent_sessions_traces,
  title   = {Real Pi Coding Agent Traces Sessions},
  author  = {Maximus Adeola},
  year    = {2026},
  url     = {https://huggingface.co/datasets/MaxDevv/real-pi-coding-agent-traces-sessions},
  note    = {Compilation of 21 pi-share-hf session datasets; see manifest.jsonl for per-session attribution}
}

Please also credit the upstream authors (see Sources table) — they did the actual work and chose to publish their traces.

Acknowledgments & disclaimer

Thanks to everyone publishing sessions via pi-share-hf, and to Mario Zechner (badlogic) for pi and the pi-share-hf pipeline that makes this kind of dataset possible.

This is an unofficial community aggregation. It is not affiliated with, endorsed by, or sponsored by pi, its authors, or any of the source dataset authors. All trademarks and project names belong to their respective owners.

Downloads last month
172