Downloads · 30 days
3
7% of all-time downloads
imdatta0/qwen3-4b-swegym-moto-kl02-sft20k-multienriched-capped-v1-adapter
qwen3-4b-swegym-moto-kl02-sft20k-multienriched-capped-v1-adapter is a text generation model from imdatta0. Use it when you need the model to write or continue text. It is set up for peft.
LoRA adapter for unsloth/Qwen3-4B-Instruct-2507, trained for SWE-Gym Moto patch generation with the search/replace edit contract.
Downloads · 30 days
3
7% of all-time downloads
All-time downloads
45
Public
Repo size
144 MB
Likes
0
Public
Click a slice to open those files.
.safetensors132 MB · 92%
From the Hugging Face model README
LoRA adapter for unsloth/Qwen3-4B-Instruct-2507, trained for SWE-Gym Moto patch generation with the search/replace edit contract.
This is a durability upload, not the promoted best 4B artifact. The current best 4B checkpoint remains the hard-multi 20k SFT adapter:
imdatta0/qwen3-4b-swegym-moto-kl02-sft20k-hardmulti-adapter.
/mnt/disks/unslothai/datta0/cache/qwen3-grpo-patch/20260605_093333_swegym_q4b-kl02-sft20k-multienriched-capped-v1_ee83de6/checkpoints/best_holdout
runs/distill_multi_enriched_capped_v1/sft_pass.jsonl.Evaluation summary: runs/passk_q4b_kl02_sft20k_multienriched_capped_v1_20k_bf16_seed9012/summary.json.
Settings: search/replace, anchored 20k context, top_k_files=16, bfloat16, max sequence 12288, structural multi-file hint, symbol-aware retrieval, ast_bonus=25, seed 9012, k=8, temperature 0.8, top_p 0.95.
| split | greedy | selected@1 | pass@8 |
|---|---|---|---|
| overall | 8/35 | 8/35 | 11/35 |
| single-file | 8/18 | - | 9/18 |
| multi-file | 0/17 | - | 2/17 |
Compared with retention-capped replay, this recovers one overall pass@8 task and one multi-file pass@8 task. It remains below the hard-multi 20k frontier, which measured 16/35 overall pass@8 and 4/17 multi-file pass@8 on seed 9012.