datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
cc-traces-weka-062126
semianalysisai/cc-traces-weka-062126
WekaTrace corpus derived from SemiAnalysis Claude Code proxy traces. Built 2026-06-21 17:48:24 UTC via utils/agentic/build_weka_hf_dataset.py.
Filters
Trace version: exactly v7
min Anthropic requests per session: 20
Claude Code CLI ≥ 2.1.139 (every row)
peak concurrent sub-agent groups ≤ 10
Non-image rows only (image content excluded at source)
Classifier calls excluded (max_tokens<=64 AND no tools → SUGGESTION MODE, title-gen… See the full description on the dataset page: https://huggingface.co/datasets/semianalysisai/cc-traces-weka-062126.cc-traces-weka-062126-256k
semianalysisai/cc-traces-weka-062126-256k
WekaTrace corpus derived from SemiAnalysis Claude Code proxy traces. Built 2026-06-21 17:49:45 UTC via utils/agentic/build_weka_hf_dataset.py.
Derived from semianalysisai/cc-traces-weka-062126 by applying the 256k per-request cap and preserving the surviving requests' relative timestamps.
Filters
Trace version: exactly v7
min Anthropic requests per session: 20
Claude Code CLI ≥ 2.1.139 (every row)
peak concurrent… See the full description on the dataset page: https://huggingface.co/datasets/semianalysisai/cc-traces-weka-062126-256k.kernelbench-mega-traces
KernelBench-Mega agent traces
Coding agents writing full GPU megakernels across Blackwell / H100 / B200, scored as speedup over reference; contamination-audited (23 verified cells).
Each .jsonl file is one agent run in Claude-Code session format, viewable with the agent trace viewer. Filename = run id; manifest.csv maps each run to model / harness / problem / GPU / score.
23 agent traces · live leaderboard: https://kernelbench.com/mega
Secrets redacted. Full reasoning for… See the full description on the dataset page: https://huggingface.co/datasets/Infatoshi/kernelbench-mega-traces.kernelbench-hard-traces
KernelBench-Hard agent traces
Frontier coding agents writing optimized CUDA/Triton kernels (FP8 GEMM, paged
attention, MoE, W4A16, KDA, Top-k) on RTX PRO 6000 Blackwell, H100 PCIe, and
B200; roofline-graded.
Each .jsonl file is one agent run in Claude-Code session format, viewable with
the Hugging Face Agent Trace viewer (Data Studio → open a row). Filename =
run id.
Live leaderboard: https://kernelbench.com/hard
Secrets redacted. Full reasoning for open-provider routes… See the full description on the dataset page: https://huggingface.co/datasets/Infatoshi/kernelbench-hard-traces.mimo-claude-code-traces-1k
MIMO Claude Code Traces
MIMO Claude Code Traces is a collection of coding-agent trajectories in a Claude Code-style environment. Each record contains a user coding task, the full multi-turn message trace, available tool schemas, assistant reasoning fields, tool calls, tool outputs, and metadata such as model name, category, duration, cost, token usage, and whether the trace used tools.
The traces were generated with mimo-v2.5-pro, MiMo's most capable model at the time of… See the full description on the dataset page: https://huggingface.co/datasets/choucsan/mimo-claude-code-traces-1k.kernelbench-cuda-tracesFable-5-traces
Glint Research Dataset Card
Fable 5 Pi Agent Traces
A compact, high-signal corpus of Fable 5 coding-agent traces converted into Hugging Face Agent Traces / Pi-compatible sessions for Data Studio inspection, tool-use policy learning, and reasoning/action distillation.
Primary Config
pi_agent/train
Agent Trace preview enabled
4,665 Pi trace sessions
60 source sessions
3,799 tool… See the full description on the dataset page: https://huggingface.co/datasets/PRATHAM4567/Fable-5-traces.paper-sim-n128-traces
WARP-RM paper simulation n=128 traces
Canonical paper-A rollout traces for the two released simulation policies over
the fixed 128-episode evaluation suite. They are the original trace pair used
by the public evaluator self-test, rather than a later re-run.
The archive contains 256 qpos traces: 128 vanilla-policy traces and 128
WARP-RM-policy traces. It is deliberately limited to evaluation output (no
model weights, DINO features, or raw source-capture data).
With the released… See the full description on the dataset page: https://huggingface.co/datasets/uynitsuj/paper-sim-n128-traces.swe-bench-lite-agent-traces-v14
AgentBRANE SWE-bench Lite Agent Traces v14
This release contains the 1,890 harness-native agent traces selected by the sealed SWE-bench Lite v14 publication record (1,379/1,890 resolved, 73.0%). It includes Claude Code, Codex, and Pi sessions across seven models and three replicates. No internal research notes are included.
Load the observation table:
from datasets import load_dataset
traces = load_dataset("melissapan/swe-bench-lite-agent-traces-v14", split="train")
Each row… See the full description on the dataset page: https://huggingface.co/datasets/melissapan/swe-bench-lite-agent-traces-v14.spec_cpu_branch_tracesFable-5-traces
Glint Research Dataset Card
Fable 5 Pi Agent Traces
A compact, high-signal corpus of Fable 5 coding-agent traces converted into Hugging Face Agent Traces / Pi-compatible sessions for Data Studio inspection, tool-use policy learning, and reasoning/action distillation.
Primary Config
pi_agent/train
Agent Trace preview enabled
4,665 Pi trace sessions
60 source sessions
3,799 tool… See the full description on the dataset page: https://huggingface.co/datasets/shijunhao/Fable-5-traces.optiq-code-traces
OptiQ Code Traces
Gold-verified agentic software-engineering trajectories, produced by OptiQ Code, the terminal coding agent for local models on a Mac. Each trajectory is a full tool-calling run against a real repository bug, and every resolved label is set by executing the gold tests (FAIL_TO_PASS + PASS_TO_PASS) after applying the model's patch, never by the agent's own self-report.
The dataset is 1,789 agent sessions in HuggingFace Session-Traces format (the agent-traces… See the full description on the dataset page: https://huggingface.co/datasets/mlx-community/optiq-code-traces.fable-5-claude-code-traces
Fable 5 Claude Code Traces
A full, scrubbed release of Fable 5 Claude Code session traces for researchers studying real coding-agent behavior: multi-turn prompts, assistant responses, tool calls, command output, retries, and session-level workflow metadata.
This release keeps the full package intact: 18 sessions, 9,497 JSONL events, 0 excluded rows, and 0 quarantine files. The traces are preserved in the native agent-event format so they can be inspected in Hugging Face Agent… See the full description on the dataset page: https://huggingface.co/datasets/AlinCiocan/fable-5-claude-code-traces.Fable-5-traces
Glint Research Dataset Card
Fable 5 Pi Agent Traces
A compact, high-signal corpus of Fable 5 coding-agent traces converted into Hugging Face Agent Traces / Pi-compatible sessions for Data Studio inspection, tool-use policy learning, and reasoning/action distillation.
Primary Config
pi_agent/train
Agent Trace preview enabled
4,665 Pi trace sessions
60 source sessions
3,799 tool… See the full description on the dataset page: https://huggingface.co/datasets/sornnakub/Fable-5-traces.codex_swebenchpro_tracesThis is a dataset generated by real swebenchpro agentic workload trace + codex agent.
1. Eval Result Summary
Metric
Value
Total trials
731
Successful trials
610
Failed trials
120
No data (skipped)
1
Passed
329
Pass rate (of successful)
53.9%
Per-Repo Breakdown
Repo
Total
Success
Failed
Passed
Pass%
ansible/ansible
96
93
3
60
65%
internetarchive/openli
91
88
3
52
59%
flipt-io/flipt85
82
3
26
32%
qutebrowser/qutebrowse
79
78
1… See the full description on the dataset page: https://huggingface.co/datasets/Inferact/codex_swebenchpro_traces.Fable-5-traces
Glint Research Dataset Card
Fable 5 Pi Agent Traces
A compact, high-signal corpus of Fable 5 coding-agent traces converted into Hugging Face Agent Traces / Pi-compatible sessions for Data Studio inspection, tool-use policy learning, and reasoning/action distillation.
Primary Config
pi_agent/train
Agent Trace preview enabled
4,665 Pi trace sessions
60 source sessions
3,799 tool… See the full description on the dataset page: https://huggingface.co/datasets/kira/Fable-5-traces.zeus-30k-trace-dataset
Zeus Qwen execution dataset
Exactly 30,000 unique solver-input traces generated through the user-provided Qwen3.8-27B API. The endpoint reports this model ID; weights were not independently inspected. Thirty concurrent API workers were requested, with fallback to twenty after repeated overload errors.
Categories: {"clarification":2000,"coding":6000,"documents":5000,"failure_recovery":2000,"maths":6000,"tool_use":4000,"writing":5000}.
Contents
traces:… See the full description on the dataset page: https://huggingface.co/datasets/Flownium/zeus-30k-trace-dataset.test-traces
Test Traces
Codex-style rollout JSONL traces exported from fast-agent for validation against the Hugging Face Agent Trace Viewer.
Fable-5-traces
Glint Research Dataset Card
Fable 5 Pi Agent Traces
A compact, high-signal corpus of Fable 5 coding-agent traces converted into Hugging Face Agent Traces / Pi-compatible sessions for Data Studio inspection, tool-use policy learning, and reasoning/action distillation.
Primary Config
pi_agent/train
Agent Trace preview enabled
4,665 Pi trace sessions
60 source sessions
3,799 tool… See the full description on the dataset page: https://huggingface.co/datasets/ApertureQA/Fable-5-traces.Fable-5-tracesA simple dataset of the raw Fable 5 Claude session logs we could get our hands on before it was taken away (no clue if it's coming back).
The raw trace files live in sessions/*.jsonl. Cache files, paste-cache files, shell history, and merged COT training exports are intentionally omitted so Hugging Face Datasets can load the repo through the agent-traces path.
A pretty viewer for dataset:… See the full description on the dataset page: https://huggingface.co/datasets/cfahlgren1/Fable-5-traces.minimax-m3-claude-code-tracesThis dataset was generated using teich by TeichAI
Prepare these datasets for supervised fine-tuning in just a few lines of code — see the Conversion section below.
Minimax M3 Claude Code Traces
This directory contains raw agent trace files generated by teich.
All assistant responses were generated by minimax/minimax-m3.
JSONL files: 31
Format
Each file is newline-delimited JSON representing a single captured agent session.
The trace schema is designed for… See the full description on the dataset page: https://huggingface.co/datasets/armand0e/minimax-m3-claude-code-traces.qwen3.7-max-pi-tracesThis dataset was generated using teich by TeichAI
Prepare these datasets for supervised fine-tuning in just a few lines of code — see the Conversion section below.
Qwen3.7 Max Pi Traces
This directory contains raw agent trace files generated by teich.
All assistant responses were generated by qwen/qwen3.7-max.
JSONL files: 47
Training-ready tools
A complete configured tools schema snapshot is embedded in the collapsed section at the bottom of this README.
Use it… See the full description on the dataset page: https://huggingface.co/datasets/armand0e/qwen3.7-max-pi-traces.kimi-k2.6-claude-code-tracesThis dataset was generated using teich by TeichAI
Prepare these datasets for supervised fine-tuning in just a few lines of code — see the Conversion section below.
Kimi K2.6 Claude Code Traces
This directory contains raw agent trace files generated by teich.
All assistant responses were generated by moonshotai/kimi-k2.6.
JSONL files: 36
Format
Each file is newline-delimited JSON representing a single captured agent session.
The trace schema is designed for… See the full description on the dataset page: https://huggingface.co/datasets/armand0e/kimi-k2.6-claude-code-traces.trace-cheating-recall-500
Trace Cheating Recall 500
This dataset contains 500 SWE-agent traces selected to evaluate whether an LLM judge
detects observable solution leakage. It is the public data source for the
trace-cheating-recall-500 Prime environment.
The examples were derived from
PrimeIntellect/int4-syn-gen-swe-glm53-bash-2026-09-02
at revision 0e7a9ecddce8de9ea8f8c369b2dd39411d6dee7a.
Composition
500 unique traces, all labeled CHEATING
250 internet-retrieval cases
250 Git-history… See the full description on the dataset page: https://huggingface.co/datasets/PrimeIntellect/trace-cheating-recall-500.Fable-5-traces
Glint Research Dataset Card
Fable 5 Pi Agent Traces
A compact, high-signal corpus of Fable 5 coding-agent traces converted into Hugging Face Agent Traces / Pi-compatible sessions for Data Studio inspection, tool-use policy learning, and reasoning/action distillation.
Primary Config
pi_agent/train
Agent Trace preview enabled
4,665 Pi trace sessions
60 source sessions
3,799 tool… See the full description on the dataset page: https://huggingface.co/datasets/edbuildingstuff/Fable-5-traces.hf-coding-tools-traces_april26
HuggingFace AI Coding Tools — Agent Traces
This dataset rehydrates the benchmark results from
davidkling/hf-coding-tools-dashboard
into the JSONL session format consumed by the
Hugging Face Agent Trace Viewer.
What's inside
32 sessions, one per (tool, model, effort, thinking) configuration
9,130 query → response turns total (≈18,260 events)
Tools covered: claude_code, codex, copilot, cursor
Models: claude-opus-4-6, claude-sonnet-4-6, claude-sonnet-4.6, composer-2… See the full description on the dataset page: https://huggingface.co/datasets/clem/hf-coding-tools-traces_april26.trace-bench
TRACE
Temporal Audit and Condition-aware Evaluation of Streaming Video Understanding
Timestamped video QA and proactive-response annotations for evaluating what a model knows, when it knows it, and how it responds.
How TRACE works
TRACE separates causal video delivery, model interaction, and scoring. The same
public contract makes QA and Proactive Response results auditable across models:
QA: answer a question using only the video evidence… See the full description on the dataset page: https://huggingface.co/datasets/omlab/trace-bench.python-mental-execution-traces
Python Mental Execution Traces
A 12,000-row prompt/completion dataset for evaluating and training language models to mentally execute self-contained Python 3 snippets without running them. Completions provide the expected standard output together with a concise variable trace or explanation.
Dataset structure
The JSONL file contains two text fields:
prompt: a Python mental-execution problem.
completion: the expected stdout and concise reasoning or variable trace.… See the full description on the dataset page: https://huggingface.co/datasets/ILoveBuns/python-mental-execution-traces.hf-coding-tools-traces-run-april12
HuggingFace AI Coding Tools — Agent Traces
This dataset rehydrates the benchmark results from
davidkling/hf-coding-tools-dashboard
into the JSONL session format consumed by the
Hugging Face Agent Trace Viewer.
What's inside
31 sessions, one per (tool, model, effort, thinking) configuration
8,875 query → response turns total (≈17,750 events)
Tools covered: claude_code, codex, copilot, cursor
Models: claude-opus-4-6, claude-sonnet-4-6, claude-sonnet-4.6, composer-2… See the full description on the dataset page: https://huggingface.co/datasets/davidkling/hf-coding-tools-traces-run-april12.hf-hub-session-pi-traces
dacorvo/hf-hub-session-pi-traces
pi coding-agent session traces produced by
agentcap runs. Each run
contributes one folder under data/<run_id>/; inside, one file per
session in pi's native export format.
The on-the-wire HTTP captures for these same runs live in
dacorvo/hf-hub-session-captures.
Both belong to the
hf-hub-session Collection
— join on run_id to align captures with traces.