OPSD-Qwen3.5-9B-LoRA-iter32

Qwen3.5-9B OPSD LoRA, iteration 32

This repository contains the adapter-only snapshot produced after OPSD iteration 32. The frozen Qwen3.5-9B base model is intentionally excluded.

Important format note

adapter/checkpoint_iter_0000032.safetensors is a raw Slime/Megatron adapter, not a PEFT directory. It contains 168 bf16 tensors whose names end in .lora_A or .lora_B. Load it through the bundled Qwen3.5 Slime model plugin with rank 64, alpha 128, dropout 0, and scaling 2.0. PeftModel.from_pretrained is not compatible with this file.

Verify the adapter before use:

sha256sum adapter/checkpoint_iter_0000032.safetensors
# 81c22bdc1d9328bb16c97483017f0570ed990cca61ec26b438d7a96e5b7557f7

Contents

• adapter/: iteration-32 LoRA and its machine-readable manifest.
• config/training_config.json: the effective runtime configuration, not the stale defaults in the original sbatch file.
• evaluation/: the 2,381-row SIF-backed SWE-Gym JSONL and eval-only launcher.
• source/: git patches, repository revisions, and untracked experiment files needed to reconstruct the dirty training tree.
• docs/: experiment and adapter-format documentation.

Evaluation

The target server must already have the same Qwen3.5-9B base model, Slime, Megatron-LM, SGLang, mini-swe-agent, Apptainer, and enough local space for the SWE-Gym SIF cache. Restore the source overlay described in source/README.md, then run:

cd /path/to/OPSD_baseline/slime

NUM_GPUS=8 \
TENSOR_MODEL_PARALLEL_SIZE=8 \
ROLLOUT_NUM_GPUS_PER_ENGINE=4 \
LOCAL=/fast/local/opsd-eval \
bash examples/coding_agent_rl/evaluate_qwen35_9b_opsd_lora.sh \
  /path/to/repo/adapter/checkpoint_iter_0000032.safetensors \
  /path/to/Qwen3.5-9B \
  /path/to/repo/evaluation/swe-gym-opsd.jsonl

The default effective sandbox concurrency is 32 with one trajectory per problem. This is not pass@32. Set N_SAMPLES_PER_EVAL_PROMPT explicitly only when repeated sampling per problem is intended.

Results are written below RUN_ROOT, including run.log and rollout_dumps/rollout_eval_0.pt. The aggregate metric is logged as eval/swegym.

The bundled JSONL is the training split, so its score measures in-distribution behavior and is not an unbiased held-out generalization result. Use a separately prepared validation/test JSONL for final reporting.

The JSONL also contains host-side SWE-Gym grading metadata and privileged golden patches. The sandbox isolation code keeps those fields hidden from the student during rollout, but they are intentionally present in this public reproduction bundle.

Excluded artifacts

This repository does not include the 9B base model, full Megatron checkpoint, Adam moments, Ray state, SIF images, or W&B credentials.