Datasets:

nanoswe
/

swesmith-qwen3.6-35b-a3b

messages listlengths 19 395	patch stringlengths 58 4.21M	model stringclasses 1 value	instance_id stringlengths 30 83	verified bool 1 class	resolved bool 2 classes	graded bool 2 classes
[ { "content": "You are a helpful assistant that can interact multiple times with a computer shell to solve programming tasks.\nYour response must contain exactly ONE bash code block with ONE command (or commands connected with && or \|\|).\n\nInclude a THOUGHT section before your command where you explain your rea...	diff --git a/check_record_created.py b/check_record_created.py new file mode 100644 index 0000000..54a89d5 --- /dev/null +++ b/check_record_created.py @@ -0,0 +1,17 @@ +import logging +import time +from datetime import datetime, timezone + +# Create a custom handler to inspect the record +class InspectHandler(logging.H...	harbor-v2	madzak__python-json-logger.5f85723f.lm_rewrite__cb3vrzvy__s0	false	true	true
[ { "content": "You are a helpful assistant that can interact multiple times with a computer shell to solve programming tasks.\nYour response must contain exactly ONE bash code block with ONE command (or commands connected with && or \|\|).\n\nInclude a THOUGHT section before your command where you explain your rea...	diff --git a/final_verification.py b/final_verification.py new file mode 100644 index 0000000..6332201 --- /dev/null +++ b/final_verification.py @@ -0,0 +1,59 @@ +""" +Final verification of the fix for all_fn and none_fn functions. +This script verifies that the bug described in the PR has been fixed. +""" +from funcy ...	harbor-v2	Suor__funcy.207a7810.combine_file__be6tat1i__s0	false	true	true
[ { "content": "You are a helpful assistant that can interact multiple times with a computer shell to solve programming tasks.\nYour response must contain exactly ONE bash code block with ONE command (or commands connected with && or \|\|).\n\nInclude a THOUGHT section before your command where you explain your rea...	diff --git a/CHANGES_SUMMARY.md b/CHANGES_SUMMARY.md new file mode 100644 index 0000000..8d46cb5 --- /dev/null +++ b/CHANGES_SUMMARY.md @@ -0,0 +1,36 @@ +# Summary of Changes to Fix Debug Module Issues + +## Problem +The debug module's function signature representation was producing incorrect output: +1. Function names...	harbor-v2	Suor__funcy.207a7810.func_basic__17s3myr6__s0	false	true	true
[{"content":"You are a helpful assistant that can interact multiple times with a computer shell to s(...TRUNCATED)	"diff --git a/funcy/seqs.py b/funcy/seqs.py\nindex 7ece4bd..81cc2bf 100644\n--- a/funcy/seqs.py\n+++(...TRUNCATED)	harbor-v2	Suor__funcy.207a7810.combine_file__rdclwp0z__s0	false	false	true
[{"content":"You are a helpful assistant that can interact multiple times with a computer shell to s(...TRUNCATED)	"diff --git a/final_test.py b/final_test.py\nnew file mode 100644\nindex 0000000..1086ca9\n--- /dev/(...TRUNCATED)	harbor-v2	seatgeek__thefuzz.8a05a3ee.combine_file__usawjinw__s0	false	false	true
[{"content":"You are a helpful assistant that can interact multiple times with a computer shell to s(...TRUNCATED)	"diff --git a/dataset/table.py b/dataset/table.py\nindex 7a3c18e..874729b 100644\n--- a/dataset/tabl(...TRUNCATED)	harbor-v2	pudo__dataset.5c2dc8d3.func_pm_ctrl_invert_if__pemmo1zy__s0	false	false	true
[{"content":"You are a helpful assistant that can interact multiple times with a computer shell to s(...TRUNCATED)	"diff --git a/final_verification.py b/final_verification.py\nnew file mode 100644\nindex 0000000..c0(...TRUNCATED)	harbor-v2	madzak__python-json-logger.5f85723f.func_pm_ctrl_shuffle__ypab5aii__s0	false	false	true
[{"content":"You are a helpful assistant that can interact multiple times with a computer shell to s(...TRUNCATED)	"diff --git a/final_verification.py b/final_verification.py\nnew file mode 100644\nindex 0000000..56(...TRUNCATED)	harbor-v2	seatgeek__thefuzz.8a05a3ee.func_basic__zk80r0js__s0	false	true	true
[{"content":"You are a helpful assistant that can interact multiple times with a computer shell to s(...TRUNCATED)	"diff --git a/reproduce_issue.py b/reproduce_issue.py\nnew file mode 100644\nindex 0000000..6112821\(...TRUNCATED)	harbor-v2	seatgeek__thefuzz.8a05a3ee.lm_rewrite__718cg7rd__s0	false	false	true
[{"content":"You are a helpful assistant that can interact multiple times with a computer shell to s(...TRUNCATED)	"diff --git a/final_verification.py b/final_verification.py\nnew file mode 100644\nindex 0000000..19(...TRUNCATED)	harbor-v2	Suor__funcy.207a7810.combine_file__kvdn5bdb__s0	false	false	true

End of preview. Expand in Data Studio

SWE-smith trajectories from Qwen3.6-35B-A3B

Multi-turn coding-agent trajectories (issue → tool-using rollout → patch) produced by Qwen3.6-35B-A3B on SWE-smith tasks, stored untokenized. This is the exact SFT corpus used for the harbor arm of the nanoswe teacher-distillation experiments.

101,901 trajectories over 45,242 unique SWE-smith task instances (3 sampled rollouts per task, ~2.25 surviving filtering), 53 parquet shards, ~1.4 GB.
≈1.96B training tokens = exactly one epoch (measured packed, SmolLM3 tokenizer, chat-rendered; ~72 messages and ~69k characters per trajectory).
51.2% of graded trajectories resolve the task — see resolved / graded below.

Schema

column	type	description
`messages`	`list<struct<role, content>>`	the trajectory as a chat conversation: `system`, `user` (task), then alternating `assistant`/`user` (observation), ending on `assistant`
`patch`	`string`	the final diff the trajectory produced (non-empty for every row)
`model`	`string`	generator tag — constant `harbor-v2`
`instance_id`	`string`	`<swesmith_instance_id>__s{0,1,2}`; the `__sN` suffix is the sample index, so the three rollouts of one task share a prefix
`graded`	`bool`	the patch was actually run against the SWE-smith test harness
`resolved`	`bool`	the patch passed the harness (F2P + P2P). Only meaningful where `graded == True`
`verified`	`bool`	legacy, constant `False` — kept only for schema compatibility with `nanoswe/nanoswe-trajs-260812`. Use `resolved`.

Correctness

resolved == False is not the same as "failed" — 9.3% of rows were never graded (the grading sweep did not cover every rollout). Always pair the two flags:

	rows	share
graded & resolved	47,273	46.4%
graded & failed	45,105	44.3%
ungraded (label unknown)	9,523	9.3%
total	101,901

Per sample:

sample	rows	graded	resolved (of graded)
`__s0`	34,065	32,878 (96.5%)	16,336 (49.7%)
`__s1`	34,036	29,739 (87.4%)	15,430 (51.9%)
`__s2`	33,800	29,761 (88.1%)	15,507 (52.1%)
all	101,901	92,378 (90.7%)	47,273 (51.2%)

Failed rollouts are deliberately kept: they are useful for RFT/DPO-style setups and for studying failure modes. Filter on resolved if you only want successful ones.

Generation


Policy	Qwen3.6-35B-A3B, SFT'd on a filtered SWE-smith curation before rollout (not the stock instruct model)
Scaffold	mini-swe-agent, bash-only, `THOUGHT:` + ```bash action format
Sampling	temperature 0.7, thinking disabled, 3 independent samples per task
Tasks	SWE-smith `smith_harbor` train split, instances 0–56,638
Serving	vLLM, per-instance Apptainer sandboxes on an HTCondor cluster

Filtering

Applied in order, from raw rollouts to the shards here:

Submitted only — rollouts that terminated without producing a patch are dropped.
Patch hygiene — cheat-pattern rollouts (e.g. recovering the fix from git history rather than writing it) are removed at consolidation.
32k fit — trajectories whose full chat render exceeds 32,768 tokens are dropped (Qwen3 tokenizer), so every row trains at a 32k sequence length without truncation.

The system message is a fixed scaffold prompt, identical across rows; downstream nanoswe training strips it and unwraps the <pr_description> envelope in messages[1].

Caveats

Task-level correlation. Up to 3 rollouts share a task instance. Split on the instance_id prefix (strip __sN), not on rows, or held-out sets will leak.
resolved is trajectory-level, not instance-level: a task can be resolved by one sample and failed by another. That disagreement is signal, not noise.
Teacher is fine-tuned, so this corpus reflects an already SWE-adapted policy — it is not a measurement of stock Qwen3.6-35B-A3B's agentic ability.
SWE-smith tasks are synthetic (bugs injected into real repos), so patch style is narrower than real-world PRs.

Loading

from datasets import load_dataset

ds = load_dataset("nanoswe/swesmith-qwen3.6-35b-a3b", split="train")
correct = ds.filter(lambda r: r["resolved"])                 # 47,273 rows
failed  = ds.filter(lambda r: r["graded"] and not r["resolved"])  # 45,105 rows

nanoswe/nanoswe-trajs-260812 — the larger 1.58M-trajectory mixed-origin corpus used for nanoswe pretraining.
nanoswe — the speedrun training code.

Downloads last month: 202