Datasetpublic

nanoswe/nanoswe-trajs-260812

nanoswe SWE-agent trajectories (v0) A consolidation of the SWE-bench-style coding-agent trajectory corpora used to train the nanoswe speedrun models. Each row is one multi-turn agent trajectory (issue → tool-using rollout → patch), stored untokenized. 1,582,701 trajectories, 34 parquet shards, content-deduplicated on traj_hash. Seed corpus: ricdomolm/mini-coder-trajs-400k; the rest are derived SWE-smith / openhands / swe-zero conversions. Schema column… See the full description on the dataset page: https://huggingface.co/datasets/nanoswe/nanoswe-trajs-260812.

sourceHugging Facemitupdated 1mo agoView on Hugging Face

0likes470downloads

Dataset Card

nanoswe SWE-agent trajectories (v0)

A consolidation of the SWE-bench-style coding-agent trajectory corpora used to train the nanoswe speedrun models. Each row is one multi-turn agent trajectory (issue → tool-using rollout → patch), stored untokenized.

—1,582,701 trajectories, 34 parquet shards, content-deduplicated on traj_hash.
—Seed corpus: `ricdomolm/mini-coder-trajs-400k`; the rest are derived SWE-smith / openhands / swe-zero conversions.

Schema

column	type	description
`messages`	`list<struct<role, content>>`	the trajectory as a chat conversation
`patch`	`string`	the final diff the trajectory produced
`model`	`string`	generator model that produced the rollout
`instance_id`	`string`	task instance identifier
`verified`	`bool`	correctness/verification flag — meaning varies by `origin`, see below
`origin`	`string`	source family (see table)
`traj_hash`	`string`	md5 of `messages`; stable dedup key

Origins

`origin`	rows	`verified=True`	generator	what `verified` means here
`ricdomolm/mini-coder-trajs-400k`	387,296	83,114	mixed	trajectory was graded; `True` = resolved the task
`swe-zero`	286,044	0	openhands-converted	ungraded (no label)
`swe-zero-extra`	74,663	0	qwen3-coder-30b-sft-v2	ungraded (no label)
`swe-hero-extra`	55,255	0	qwen3-coder-30b-sft-v2	ungraded but gradable (tests exist)
`swe-smith-extra`	686,986	276,216	qwen3-coder-30b-sft-v2-rfreeze	instance-level verification (verified vs unverified task)
`swe-smith-extra-short`	92,457	35,242	qwen3coder-30b-brief	trajectory was graded per-row

⚠️ `verified` is not homogeneous

verified=False means different things depending on origin:

—graded and failed — ricdomolm/mini-coder-trajs-400k
—never graded (no label) — swe-zero, swe-zero-extra, swe-hero-extra
—task instance not verified — swe-smith-extra

Use origin to disambiguate. The model column further separates the ungraded bulk generators (qwen3-coder-30b-sft-v2*, openhands-converted) from graded sets. Treat verified=True as a positive signal only within an origin that was actually graded.

Provenance notes

—swe-zero is the intersection of two renderings (str_replace/fit32k and render32k) of the same rollouts, keeping the render32k messages.
—ricdomolm/mini-coder-trajs-400k here is the union of the correct + incorrect-split1 + incorrect-split2 trajectory sets.
—"render32k" = trajectory whose nanoswe render is ≤32k tokens; "trunc100" = first 100 messages kept.

SWE-bench source-repo exclusion

To keep this corpus clean for SWE-bench-style evaluation, 4,516 trajectories whose source environment is one of the 12 SWE-bench source repositories were removed (decision key: the upstream owner/repo encoded in instance_id). No trajectory in this dataset is sourced from any of: astropy/astropy, django/django, matplotlib/matplotlib, mwaskom/seaborn, pallets/flask, psf/requests, pydata/xarray, pylint-dev/pylint, pytest-dev/pytest, scikit-learn/scikit-learn, sphinx-doc/sphinx, sympy/sympy.

Removed counts by repo: sympy/sympy 1,778 · pallets/flask 1,329 · pydata/xarray 676 · sphinx-doc/sphinx 361 · astropy/astropy 258 · mwaskom/seaborn 112 · scikit-learn/scikit-learn 2 (the other five repos contributed 0 rows). Ecosystem-sibling repos that merely share an org or name family (e.g. pylint-dev/astroid, pallets/jinja, pallets/werkzeug, astropy/astroquery, django/channels, the pytest-dev/pytest-* plugins) are distinct codebases, not forks of the 12, and are retained.

Loading

python

from datasets import load_dataset
ds = load_dataset("ricdomolm/nanoswe-trajs-v0", split="train")