datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Open-SWE-Traces
Open-SWE-Traces: Advancing Distillation for Software Engineering Agents
🚨 What's New
[09/26] Release v1.2: Added new agent trajectories generated by Qwen3.8-27B for
mini-swe-agent. Trajectories for OpenCode and Claude Code harnesses will be released soon.
[08/26] Release v1.1: Added new agent trajectories generated by DeepSeek-V4-Flash and
Qwen3.6-27B across OpenHands,
SWE-agent, and mini-swe-agent harnesses.
[06/21] Release v1.0: Released 207k agent… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Open-SWE-Traces.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.cot-eval-traces-2.0kernelbench-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.real-pi-coding-agent-traces-sessions
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… See the full description on the dataset page: https://huggingface.co/datasets/MaxDevv/real-pi-coding-agent-traces-sessions.gpt-oss-20b-moe-expert-power-traces-320k
GPT-OSS-20B MoE Expert Power Traces (320k, ChipWhisperer)
This dataset contains analog power traces captured with a ChipWhisperer Husky while running forced single-expert MoE computations derived from openai/gpt-oss-20b on an NVIDIA H100.
What is recorded
Each trace corresponds to one capture trial where:
A fixed expert id is selected (expert_00 ... expert_31).
A random hidden-state tensor is generated once per trial.
The selected expert computation is executed… See the full description on the dataset page: https://huggingface.co/datasets/masterpieceexternal/gpt-oss-20b-moe-expert-power-traces-320k.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.chankhavu-imo-reasoning-tracesfunes-nvidia-Open-SWE-Traces
Funes recall store — NVIDIA Open-SWE-Traces (resolved)
A funes recall store built by indexing the
resolved==1 trajectories of
nvidia/Open-SWE-Traces
(65244 sessions, across both harnesses — SWE-agent and OpenHands — and both models,
Minimax-M2.5 and Qwen3.5-122B).
What this is
This is not a raw trace dataset — it is a pre-built funes index: the source
trajectories chunked into content blocks and embedded, stored as a
Lance table (chunks.lance).
Source… See the full description on the dataset page: https://huggingface.co/datasets/dacorvo/funes-nvidia-Open-SWE-Traces.jobseek-agent-traces
Jobseek Agent Traces
Claude Code agent session traces from jobseek — a job posting monitor for company career pages.
Each trace captures a complete agent workflow session: company discovery, board configuration, monitor/scraper selection, and quality validation. These are raw session transcripts, not tabular data — use the trace viewer to explore them.
Structure
traces/
{company-slug}/
{date}.jsonl # One trace per session (header + records)
Each .jsonl… See the full description on the dataset page: https://huggingface.co/datasets/viktor-shcherb/jobseek-agent-traces.misc-merged-claude-code-traces-v1
MISC Unification of Public Claude Code Traces
A unified dataset of 32,133 deduplicated Claude API conversation traces focused on software engineering and code generation tasks. This dataset merges and normalizes traces from 10 different source datasets into a single, consistent format.
Dataset Description
This dataset contains real Claude API interaction traces capturing software engineering workflows including:
Code generation and modification
Bug fixing and debugging… See the full description on the dataset page: https://huggingface.co/datasets/nlile/misc-merged-claude-code-traces-v1.hal_tracesOx-Alpha-Pi-TracesThis dataset was generated using teich by TeichAI
Ox-Alpha Pi Agent Coding Traces
This directory contains raw agent trace files generated by teich.
JSONL files: 2247
Model metadata: stealth/ox-alpha
Domains and prompt distribution
Topic
Traces
Games & simulation (headless)
196
Frontend & Node-testable web
159
Health & medicine informatics
139
ML & scientific computing (CPU)
123
Data analysis & reporting
122
Computational biology & chemistry… See the full description on the dataset page: https://huggingface.co/datasets/TeichAI/Ox-Alpha-Pi-Traces.aime_1983_2023_deepseek-r1_traces_16384fol-traces
citation
@misc{lee2025foltraces,
title={FOL-Traces: Verified First-Order Logic Reasoning Traces at Scale},
author={Lee, Isabelle and Liaw, Sarah and Yogatama, Dani},
year={2025},
eprint={2505.14932},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2505.14932}
}
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.wmo-terminal-tasks-traces
terminal-tasks — real agent-environment traces
Computer-use agent runs in real terminal containers: bash commands and their true outputs from live task environments.
Every trace is a REAL run: an LLM agent stepping against the actual benchmark environment, with
each transition (tool call → true environment observation) recorded as OpenTelemetry GenAI spans
(traces.otel.jsonl, one span per line). Captured by
world-model-harness's
environment-capture package, which also holds the… See the full description on the dataset page: https://huggingface.co/datasets/experiential-labs/wmo-terminal-tasks-traces.qwen36-27b-length-traces
Qwen3.6-27B generation-length prediction: heads, calibrations and workloads
Artifacts for conformal length-aware LLM scheduling on Qwen/Qwen3.6-27B — predicting a
request's remaining generation length from a hidden layer during decoding, wrapping it in a
split-conformal interval, and scheduling with SRPT inside vLLM. Extends TRAIL
(Don't Stop Me Now, ICLR'25) to a hybrid-attention reasoning model.
This repo contains the derived artifacts, not the raw activations. The 3250… See the full description on the dataset page: https://huggingface.co/datasets/dungnv/qwen36-27b-length-traces.schedulerlens-tracesagent-llm-traces-v2
Exgentic Agent LLM Traces v2 — Agent Chat Only
OpenTelemetry-shaped execution traces for 10,057 agent runs across 6 benchmarks (AppWorld, SWE-bench, BrowseCompPlus, τ²-bench Airline/Retail/Telecom), filtered to the agent under test's chat-only LLM calls. This is the dataset for replay testing, behavioral analysis, or any task where you care about what the benchmarked model actually did — not the eval scaffolding around it.
This v2 release expands upon Exgentic/agent-llm-traces… See the full description on the dataset page: https://huggingface.co/datasets/Exgentic/agent-llm-traces-v2.post-train-bench-traces
PostTrainBench Sessions by Benchmark
Derived from akseljoonas/posttrainbench-sessions on 2026-04-20.
This dataset exports each source row as one viewer-compatible JSONL trace and groups traces by benchmark.
Layout
benchmarks.json: benchmark catalog and counts
benchmarks/<benchmark>/index.json: metadata index for one benchmark
benchmarks/<benchmark>/<job_id>.jsonl: one converted session trace per source row
Benchmarks
Benchmark
Sessions
aime2025
19… See the full description on the dataset page: https://huggingface.co/datasets/smolagents/post-train-bench-traces.trustworthy-biology-agents-traces
Trustworthy Biology Agents — Run Traces
Raw execution traces from 1,329 agent runs across three coding agents on three
biology benchmarks — BiomniBench-DA, BixBench, and CompBioBench. This is the scrubbed
trace bundle for the study in
manu-tej/ai-scientists; the write-up
lives in that repo's RESULTS.md.
The motivating question is not only whether an agent reaches the right answer, but
whether it behaves like a trustworthy analyst when the task is ambiguous,
under-specified, or… See the full description on the dataset page: https://huggingface.co/datasets/amanutej/trustworthy-biology-agents-traces.agent-traces
Trace Commons — Agent Traces
Trace Commons is one open, public dataset of coding-agent sessions — the
back-and-forth between a developer and an AI coding agent, including prompts,
model responses, tool calls, and command output — contributed voluntarily as an
open resource for studying, evaluating, and building on how these agents
actually work.
Every trace here was donated only from a public, open-source repository, was
anonymized on the contributor's own machine before upload… See the full description on the dataset page: https://huggingface.co/datasets/trace-commons/agent-traces.Fable-GPT-5.5-Distillation-Traces
Agent Traces Curated 2026 (v3 Merged)
A unified distillation corpus of 9,057,143 records spanning agentic
coding traces, math/code/science reasoning, tool-use trajectories, and
preference data. 8,876,012 train + 181,131 eval, stratified by source.
What this is
This is the v3 merged corpus that supersedes both v1 and v2 of this dataset.
It combines five major source groups through a unified normalization
pipeline:
Original v2 RESMP-DEV (de-fragmented, re-deduped):… See the full description on the dataset page: https://huggingface.co/datasets/RESMP-DEV/Fable-GPT-5.5-Distillation-Traces.paper-sim-n512-traces
paper-sim-n512-traces
Rollout traces for the n=512 simulated bottle-in-bin evaluation reported in
Table I of the WARP-RM CoRL rebuttal.
Two arms, 512 paired scenes each:
directory prefix
arm
bottles/scene
throughput
all-6
fullhz_vanilla_sh00..15
Vanilla BC (100% of data)
3.885
237/hr
9.4%
fullhz_paperwarp512_sh00..15
WARP-BC (31.5% kept)
4.533
290/hr
25.0%
512 traces per arm = 16 shards x 32 worlds. Each qpos_trace_NNN_sSEED.npz
holds the full qpos trajectory… See the full description on the dataset page: https://huggingface.co/datasets/uynitsuj/paper-sim-n512-traces.terminal-bench-2.1-qwen3.8-27b-traces
Terminal-Bench 2.1 traces: Qwen3.8-27B-GPTQ-4bit, xhigh / medium / low / off
Complete agent trajectories, verifier output, timing and token usage for all 89
Terminal-Bench 2.1 tasks run locally with
btbtyler09/Qwen3.8-27B-GPTQ-4bit
on 2× RTX 3090, plus the adaptive fallback reruns at lower reasoning effort.
Headline result: 62/89 (69.66%) at xhigh in a single clean pass.
Cumulative best-of across xhigh → medium → low → off fallbacks: 70/89 (78.65%).
The second number is not a… See the full description on the dataset page: https://huggingface.co/datasets/Lottolabs/terminal-bench-2.1-qwen3.8-27b-traces.history-anchor-100-traces
History Anchor 100 — Model Trajectories
*Per-(model × condition × scenario set × seed) raw outputs from the paper "History Anchors: How Prior Behavior Steers LLM Decisions Toward Unsafe Actions".*
This dataset contains the full set of model decisions that back every figure and table in the paper. Use it to:
audit a single model's behaviour scenario-by-scenario,
recompute headline metrics without re-running the (paid) API sweeps,
mine reasoning_content traces from models that expose… See the full description on the dataset page: https://huggingface.co/datasets/albertoRodriguez97/history-anchor-100-traces.terminal-bench-traces-localkimi-k3-coding-and-debugging-traces
Kimi K3 Coding, Tool Use & Instruction Following Traces
582 TRAJECTORIES · 3,956 TRAINING ROWS · 3 MB PARQUET · 72 MB JSONL
Generated by moonshiner — an open harness for
distilling verified instruction-following, tool-use, and agentic coding traces.
Behavior-preserving instruction-following, tool-use, and agent trajectories
from Kimi K3 (moonshotai/kimi-k3). The category and row-share tables
below describe the actual mix seen during training rather than assuming a… See the full description on the dataset page: https://huggingface.co/datasets/greghavens/kimi-k3-coding-and-debugging-traces.