Dataset Viewer
Auto-converted to Parquet Duplicate
instance_id
stringlengths
11
68
messages
stringlengths
2
247k
repo
stringlengths
5
63
trajectory_format
stringclasses
1 value
exit_status
stringclasses
130 values
prompt_tokens
int64
0
775k
completion_tokens
int64
0
30k
duration_sec
float64
0.21
1.56k
alibaba__fescar-382
[{"role": "system", "content": "You are a helpful assistant that interacts with a computer to solve software-engineering tasks.\n\nEvery response must contain EXACTLY ONE bash code block (triple backticks) with EXACTLY ONE command.\nBefore the bash block, include a THOUGHT section explaining your reasoning. Put ALL exp...
alibaba/fescar
mini-swe-agent-1
Submitted
15,924
507
65.162145
alibaba__fescar-382
[{"role": "system", "content": "You are a helpful assistant that interacts with a computer to solve software-engineering tasks.\n\nEvery response must contain EXACTLY ONE bash code block (triple backticks) with EXACTLY ONE command.\nBefore the bash block, include a THOUGHT section explaining your reasoning. Put ALL exp...
alibaba/fescar
mini-swe-agent-1
Submitted
39,517
1,037
74.700621
singapore__renovate-1165
[{"role": "system", "content": "You are a helpful assistant that interacts with a computer to solve software-engineering tasks.\n\nEvery response must contain EXACTLY ONE bash code block (triple backticks) with EXACTLY ONE command.\nBefore the bash block, include a THOUGHT section explaining your reasoning. Put ALL exp...
singapore/renovate
mini-swe-agent-1
Submitted
55,183
1,270
79.234879
alibaba__fescar-382
[{"role": "system", "content": "You are a helpful assistant that interacts with a computer to solve software-engineering tasks.\n\nEvery response must contain EXACTLY ONE bash code block (triple backticks) with EXACTLY ONE command.\nBefore the bash block, include a THOUGHT section explaining your reasoning. Put ALL exp...
alibaba/fescar
mini-swe-agent-1
Submitted
60,085
1,636
86.469214
systemd__mkosi-2044
[{"role": "system", "content": "You are a helpful assistant that interacts with a computer to solve software-engineering tasks.\n\nEvery response must contain EXACTLY ONE bash code block (triple backticks) with EXACTLY ONE command.\nBefore the bash block, include a THOUGHT section explaining your reasoning. Put ALL exp...
systemd/mkosi
mini-swe-agent-1
incomplete
230,233
1,743
89.576621
google__jax-21774
"[{\"role\": \"system\", \"content\": \"You are a helpful assistant that interacts with a computer t(...TRUNCATED)
google/jax
mini-swe-agent-1
Submitted
95,161
1,903
92.283068
systemd__mkosi-2044
"[{\"role\": \"system\", \"content\": \"You are a helpful assistant that interacts with a computer t(...TRUNCATED)
systemd/mkosi
mini-swe-agent-1
incomplete
203,422
1,900
92.914954
singapore__renovate-1165
"[{\"role\": \"system\", \"content\": \"You are a helpful assistant that interacts with a computer t(...TRUNCATED)
singapore/renovate
mini-swe-agent-1
Submitted
81,197
2,123
97.043871
qiskit__qiskit-terra-6658
"[{\"role\": \"system\", \"content\": \"You are a helpful assistant that interacts with a computer t(...TRUNCATED)
Qiskit/qiskit-terra
mini-swe-agent-1
Submitted
81,495
2,160
97.762351
pagerduty__go-pagerduty-448
"[{\"role\": \"system\", \"content\": \"You are a helpful assistant that interacts with a computer t(...TRUNCATED)
PagerDuty/go-pagerduty
mini-swe-agent-1
Submitted
64,899
2,279
100.137107
End of preview. Expand in Data Studio

SWE-ZERO 1B Trajectories

96,237 execution-free agentic rollouts across 32,079 pull requests in 20 programming languages from SWE-rebench V2.

Overview

Each rollout is a multi-turn trajectory where a small code model (ricdomolm/mini-coder-1.7b, Qwen3-1.7B fine-tuned on 400k mini-swe-agent trajectories) attempts to fix a real GitHub issue using only bash commands (cat, grep, sed, find, etc.) without access to language interpreters or test execution.

  • Rollouts: 96,237 (3 per PR)
  • PRs: 32,079
  • Languages: 20 (Python, JavaScript, TypeScript, Java, C++, Go, Rust, etc.)
  • Unique content tokens: ~1.1B (Qwen3-1.7B tokenizer, mean ~11.4K per rollout)
  • Submission rate: 38.6% (model produced a patch)
  • Execution-based pass@1: 6.0% (ConTree eval on 151 Python instances)
  • Execution-based pass@3: 11.3%

Columns

Column Description
instance_id SWE-rebench V2 instance ID (org__repo-pr)
messages JSON-encoded conversation (system, user, assistant turns)
repo GitHub repository (org/repo)
trajectory_format Always mini-swe-agent-1
exit_status Submitted (model produced a patch) or other
prompt_tokens Cumulative prompt tokens across all API calls
completion_tokens Total model output tokens
duration_sec Wall-clock time for the rollout

Message Format

Each rollout's messages field is a JSON list of conversation turns:

  • system: Task instructions and tool definitions
  • user: Issue description (first turn) or bash command output (subsequent turns)
  • assistant: Model's reasoning + bash command (in bash_command field)

Usage

from datasets import load_dataset

ds = load_dataset("AlienKevin/SWE-ZERO-1B-trajectories", split="train")

# Filter to submitted rollouts
submitted = ds.filter(lambda x: x["exit_status"] == "Submitted")
print(f"Submitted: {len(submitted)} / {len(ds)}")

Generation

Generated using the SWE-ZERO pipeline from marin-community/marin on the Iris TPU cluster. See issue #4666 for reproduction commands and details.

Part of marin-community/marin#4435.

Downloads last month
181