BurnyCoder/qwen2.5-coder-0.5b-swe-rl

Qwen2.5-Coder-0.5B MiniBug-RL

This repository contains the merged output of the MiniBug-RL hidden-unit-test GRPO experiment. It targets single-function Python repair; it is not a repository-scale autonomous software-engineering agent.

Documentation correction

The model, adapter, and measured results were originally published at immutable revision `5b6e22a4c6c01bec95d10e93a0fc78666eb9c543`. This later README-only correction fixes metric terminology and narrows claims to what the recorded evidence establishes. It did not rerun training or evaluation. The audited repository documentation is at GitHub commit `902ca72a46bbc55b3154a3e2f1cd5c8cff6ddde2`.

Exact lineage

• —Base: `Qwen/Qwen2.5-Coder-0.5B-Instruct` at `ea3f2471cf1b1f0db85067f1ef93848e38e88c25`
• —Source: https://github.com/BurnyCoder/llm-rl-software-engineering/commit/7027bc55baecc00fad51cbe2b8f030dca2c91c1e
• —Trainer: TRL GRPO with LoRA rank 16, alpha 32, learning rate 1e-05, seed 42
• —Selection: train chosen only on the 12-task validation split

Frozen final-test result

Observed sampled success@4 is the fraction of tasks with at least one complete repair among exactly four generated samples; it is not an unbiased pass@k estimator.

Paired hidden-test-fraction difference: -0.0042 with paired percentile-bootstrap 95% interval [-0.2917, 0.2625]. The validation learning gate, pre-specified in the producing source commit, passed. These small synthetic-split measurements should not be generalized to SWE-bench.

Frozen external HumanEvalFix result

Protocol: `bigcode/humanevalpack` at immutable dataset revision `9a41762f73a8cb23bb5811b73d5aab164efcf378`, prompt variant humanevalfixdocs-python, greedy n=1. Postprocessing reference: pinned BigCode Python harness. Paired pass@1 difference: 0.0061 with paired percentile-bootstrap 95% interval [-0.0183, 0.0366]. HumanEvalPack examples, candidate outcomes, and scores did not enter the implemented training, reward-computation, checkpoint-selection, or tuning data flows. Its pinned harness source defined the frozen external protocol; public benchmark contamination may still affect both policies. Executable candidates ran under MiniBug-RL's host-enforced 3-second wall-clock deadline around resource-limited isolated Docker, rather than the pinned BigCode Python harness's 10-second limit; treat this as a MiniBug-sandbox measurement, not a directly comparable leaderboard score. Sandbox image: sha256:a869cd1dffb8c87afad1bb1302106cb9f5cb580641c7391bb73f4ab077f140d9. Docker controls are defense in depth, not proof that this runner is safe for arbitrary hostile code. Internal parser/policy rejections and deterministic reward cache hits do not start containers. Paired percentile-bootstrap intervals describe these fixed task samples; they do not by themselves establish broad model quality.

Included evidence

• —`results.json`: candidate-level outcomes and aggregates. Its immutable historical key sampled_pass_at_4 denotes the observed four-draw statistic relabeled above.
• —`training_config.json`: resolved non-secret configuration.
• —Producing source commit: implementation used for training and evaluation.

Load the resulting model

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "BurnyCoder/qwen2.5-coder-0.5b-swe-rl"
result_revision = "5b6e22a4c6c01bec95d10e93a0fc78666eb9c543"
tokenizer = AutoTokenizer.from_pretrained(model_id, revision=result_revision)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    revision=result_revision,
    dtype="auto",
)

The separately loadable LoRA adapter and tokenizer are in adapter/. Candidate-level measurements and the resolved configuration are included.