datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
CoderForge-Preview-32B-SWE-Bench-Verified-Evaluation-trajectoriesSWE-bench_Verified_With_Annotationsswebench-verified-deepseek-v4-flash-failure-analysis
SWE-bench Verified runs & failure analysis — DeepSeek-V4-flash (local) × mini-swe-agent
Per-instance analysis of SWE-bench Verified runs of a locally-served DeepSeek-V4-flash model
driven by mini-swe-agent, graded with the official
SWE-bench harness. Each instance carries the full agent trajectory, a readable transcript, the
submitted patch, the harness test output, deterministic metrics, and a hand-verified qualitative
root-cause diagnosis.
Current numbers (resolve rates… See the full description on the dataset page: https://huggingface.co/datasets/daaain/swebench-verified-deepseek-v4-flash-failure-analysis.agent-trajectories-swe-bench-test-minus-verified
Agent Trajectories: SWE-bench Test \ Verified — Mixed Teachers (gpt-5.2 / gpt-5-mini)
Summary
Full multi-turn agent trajectories collected from the SWE-bench Test minus Verified split
(i.e., SWE-bench Test instances that are not part of SWE-bench Verified).
Intended for SFT of agent models on coding tasks.
Data Collection
Each trajectory was produced by a GT-aware lookahead agent that, at every turn:
Sampled a candidate response from both gpt-5.2 and… See the full description on the dataset page: https://huggingface.co/datasets/JetBrains-Research/agent-trajectories-swe-bench-test-minus-verified.SWEbench-Verified-eval150-u355-M2.7-orch-cost-20260923
Fixed Solo350 u355 + MiniMax-M2.7: orchestration cost study
Best observed cost tradeoff: compact coordinator decisions plus soft review at the existing hard limit (at most 12 worker turns). M2.7 metered token cost falls 59.3%, while mean solved tasks decrease from 90.00 to 87.67/150. Accuracy equivalence was not established.
This closed study contains 6 designs and 16 complete independent runs on the same 150 tasks (2400 scored task/run pairs), each with an independent audit.… See the full description on the dataset page: https://huggingface.co/datasets/CharlieLLL/SWEbench-Verified-eval150-u355-M2.7-orch-cost-20260923.SWE-bench_Verified_Small70B_normal_llama_33_70b_instruct__swe_bench_verified_mini
Inspect Dataset: 70B_normal_llama_33_70b_instruct__swe_bench_verified_mini
Dataset Information
This dataset was created using the create_inspect_dataset function from the deception_sprint package on 2025-06-17.
Model Information
Model: vllm/meta-llama/Llama-3.3-70B-Instruct
Model args: {'max_model_len': 32768, 'gpu_memory_utilization': 0.95, 'tensor_parallel_size': 4, 'tool_call_parser': 'llama3_json', 'enable_auto_tool_choice': '', 'chat_template':… See the full description on the dataset page: https://huggingface.co/datasets/jordan-taylor-aisi/70B_normal_llama_33_70b_instruct__swe_bench_verified_mini.SWEbench-Verified-eval150-M2.7-Qwen3.5-9B-orch-7arms-2repeats-w32-20260920
SWE-bench Verified eval150 — M2.7 × Qwen3.5-9B, seven arms, two repeats, 32 concurrency
Campaign 2026-09-20. 14/14 independent full150 runs audited. Complete accuracy evidence.
Evaluation mode is orch: MiniMax-M2.7 orchestrator and the specified Qwen3.5-9B worker. Training mode is labeled independently. All runs use 32 concurrent episodes, 10GiB Docker sandboxes, four TP1 workers and one TP4/EP4 coordinator. Frozen regression-gated prompts, decoding and canonical verifier match… See the full description on the dataset page: https://huggingface.co/datasets/CharlieLLL/SWEbench-Verified-eval150-M2.7-Qwen3.5-9B-orch-7arms-2repeats-w32-20260920.swe-bench-verifiedswe_doc_gen_SWE-bench_Verified_testSWE-bench_Verified_micro
SWE-bench_Verified_Micro
A subsampled version of SWE-Bench_Verified dataset with 50 samples, sampled using the original unique repo distribution from the original dataset
devstral-swe-bench-verified
Devstral SWE-bench Verified Trajectories
This dataset contains 4,000 mini-SWE-agent trajectories generated by
mistralai/Devstral-Small-2-24B-Instruct-2512 on a fixed 250-problem training split of
SWE-bench Verified. There are 16 independently seeded rollouts per problem.
Every row preserves the exact request prefix mini-SWE-agent supplied before Devstral's
first generation, the complete ordered rollout, a sanitized final patch, and termination
metadata. Official SWE-bench… See the full description on the dataset page: https://huggingface.co/datasets/suryadv/devstral-swe-bench-verified.swebench-verified-deepseek-v4-flashswe-bench-verified-raw-traces-qwen3-coder
SWE-bench Verified raw mini-SWE-agent traces
Raw mini-SWE-agent trajectories from 20250802_mini-v1.0.0_qwen3-coder-480b-a35b-instruct for SWE-bench Verified.
The raw/easy split uses exactly the 194 instance IDs from parsaidp/SWE-bench_Verified_easy. That
public dataset contains SWE-bench Verified questions and Kimi-generated answers; this
dataset uses only its instance IDs. The trajectory contents here are local
mini-SWE-agent/Qwen traces.
Files
data/full.jsonl:… See the full description on the dataset page: https://huggingface.co/datasets/parsaidp/swe-bench-verified-raw-traces-qwen3-coder.swebench-verified-top15-embeddings-8k
SWE-bench Verified — top-15 model embeddings (Qwen3-8B, 8k)
Frozen Qwen3-8B pooled state & action embeddings (4096-d) for every agent step
of the top-15 models (by SWE-bench Verified resolved rate) from
tarsur385/swev-trm-trajectories-25models. Encoder run at 8k max-model-len
(last-token pooling; tail-truncated), same pipeline as the DeepSWE embeddings.
One row per step. 287,251 steps · 7,011 trajectories · 500 tasks · 15 models.
Columns
column
type… See the full description on the dataset page: https://huggingface.co/datasets/tarsur385/swebench-verified-top15-embeddings-8k.swebench-verified-kimi-k2p6-traces
SWE-bench Verified Kimi K2.6 Reasoning Traces
This dataset contains reasoning traces generated on princeton-nlp/SWE-bench_Verified using fireworks_ai/kimi-k2p6-high with a mini-swe-agent based harness. It is intended for research and distillation of software-engineering agents.
The repository is published with three configs because each table has a different schema:
raw_trajectories: one row per SWE-bench instance with the patch, sanitized result JSON, full trajectory JSON, message… See the full description on the dataset page: https://huggingface.co/datasets/MemoryAsModality/swebench-verified-kimi-k2p6-traces.swebench-verified-claude-opus-4.7swebench-verified-deepseek-v3.2swebench-verified-mini-swe-agent-gpt41mini-test
swebench-verified-mini-swe-agent-gpt41mini-test
Harbor evaluation on swebench-verified@1.0: 0/1 resolved (0.0%), 0 errors
Dataset Info
Rows: 1
Columns: 25
Columns
Column
Type
Description
instance_id
Value('string')
Task identifier (e.g. astropy__astropy-12907)
reward
Value('float64')
Verifier reward (e.g. 0.0 or 1.0)
resolved
Value('bool')
Whether the task was resolved (reward > 0)
agent
Value('string')
Agent name used for this trial
model… See the full description on the dataset page: https://huggingface.co/datasets/reasoning-degeneration-dev/swebench-verified-mini-swe-agent-gpt41mini-test.swe-bench-verified-raw-traces-qwen3-coder
SWE-bench Verified raw mini-SWE-agent traces
Raw mini-SWE-agent trajectories from 20250802_mini-v1.0.0_qwen3-coder-480b-a35b-instruct for SWE-bench Verified.
The raw/easy split uses exactly the 194 instance IDs from
parsaidp/SWE-bench_Verified_easy. That public dataset contains SWE-bench
Verified questions and Kimi-generated answers; this dataset uses only its
instance IDs. The trajectory contents here are local mini-SWE-agent/Qwen traces.
Files
data/full.jsonl:… See the full description on the dataset page: https://huggingface.co/datasets/nikitamounier/swe-bench-verified-raw-traces-qwen3-coder.swebench-verified-gpt-5.2swebench-verified-gemini-3-proswebench-verified-gpt-5-nanoswebench-verified-gpt-5-miniswebench-verified-gemini-2.5-proswebench-verified-trajectory-embeddings-qwen3.6-27b
SWE-bench Verified — Qwen3.6-27B trajectory embeddings (10 models, tail-truncated)
Pre-computed pooled embeddings of mini-swe-agent trajectories on SWE-bench Verified, using
Qwen/Qwen3.6-27B in vLLM runner="pooling" mode (last-token pooling).
Raw trajectories live at:
tarsur385/swebench-verified-trajectories.
Embedding recipe
Backbone: Qwen/Qwen3.6-27B
vLLM mode: runner="pooling", last-token pooling, dtype bfloat16
Chat template: the model's default; for Qwen3.6… See the full description on the dataset page: https://huggingface.co/datasets/tarsur385/swebench-verified-trajectory-embeddings-qwen3.6-27b.SWE-bench_Verified_ratedswebench-verified-gemini-3.1-pro-preview