dpsk_untitledagent_data_lanzo

Repository-repair teacher trajectories generated with deepseek/deepseek-v4.1-flash (low reasoning effort) through an API. This release contains 332 attempts across 305 unique tasks, including 200 test-verified successful attempts and 132 unsuccessful attempts. The latest completed outcome is successful for 200 tasks and unsuccessful for 105 tasks. Retries are separate attempts, not independent tasks.

Files and use

• attempts.jsonl.gz: one record per attempt; source dataset/revision, repository, base commit, issue, generated patch, evaluation outcome, duration, and test output.
• requests.jsonl.gz: all 6,781 recorded successful API responses and their model-visible input histories and tool schemas. Returned reasoning fields are preserved, including provider duplicates. API requests that failed without a model response are not trajectories and are not included.
• train.jsonl / validation.jsonl: 2,951 / 216 per-decision examples, respectively. Exact model-visible history is the prompt; target is the current assistant response. Reasoning is normalized to a single reasoning_content field. Only the current target should receive training loss.
• splits.json: deterministic task-disjoint split, 180 training tasks and 20 validation tasks. All retries of a task belong together. Repositories can overlap between splits; this is not a repository-held-out benchmark.
• sources.json: pinned upstream revisions and attribution.
• statistics.json: machine-readable counts.

Read without automatic nested-schema inference:

import gzip, json
from huggingface_hub import hf_hub_download
p = hf_hub_download('Lanzo-T-H/dpsk_untitledagent_data_lanzo',
                   'requests.jsonl.gz', repo_type='dataset')
with gzip.open(p, 'rt', encoding='utf-8') as f:
    request = json.loads(next(f))

Collection and quality limits

The agent used a repository sandbox, issue description, file/edit/shell tools, a 15-minute limit, and up to 56 tool calls. Reference patches were used by the evaluator, not intentionally supplied to the agent. Environment acceptance required the reference patch to pass and the empty patch to fail. Tool commands ran in a chroot with dropped capabilities. This is a practical isolation boundary, not a proof against all possible information leakage.

An audit of all 332 attempts found no successful reference-patch retrieval. Recorded Git history outputs had baseline commit titles only. One unsuccessful attempt tried to retrieve a public upstream source file and received a DNS failure; it is retained in the raw data and excluded from the successful training subset. This audit does not prove absence of pretraining memorization, issue-text solution hints, or unrecognized side channels. Test success is not an exhaustive semantic correctness review.

Three length-truncated responses and 22 responses with no content/tool action from successful trajectories are excluded as training targets; they remain in the raw release. Prefixes reproduce runtime compaction, rather than artificially concatenating an entire trajectory. Reasoning fields are API-returned text, not a guarantee of the model's complete internal computation. Reasoning-related sampling metadata describes the client request; only the provider-supported reasoning effort was enforced by the teacher proxy.

All formal collection attempts with recorded teacher responses are included. Calibration-only and evaluation-only runs are excluded. Reference/test patches, model credentials, HTTP authentication headers, host configuration, and private evaluation files are not released. Internal harness identifiers have been replaced consistently by neutral identifiers in text, including the baseline Git tag. Consequently these are sanitized records, not byte-identical raw transport logs. Original private archives remain separate.

Upstream attribution

• SWE-rebench-V2, Nebius, CC BY 4.0.
• R2E-Gym-Subset, R2E-Gym, Apache 2.0.
• SWE-Gym, SWE-Gym, MIT.
• SWE-rebench, Nebius, CC BY 4.0.

Each attempt identifies its source repository. Repository code excerpts and patches remain subject to their respective upstream licenses. See LICENSE.md; this release does not relicense third-party material.