Downloads · 30 days
3
7% of all-time downloads
imdatta0/qwen3-4b-swegym-moto-kl02-sft20k-hardmulti-ec2analog-capped-v1-adapter
qwen3-4b-swegym-moto-kl02-sft20k-hardmulti-ec2analog-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 checkpoint for a bounded ablation in the Qwen3 GRPO patch investigation.
Downloads · 30 days
3
7% of all-time downloads
All-time downloads
41
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 checkpoint for a bounded ablation in the Qwen3 GRPO patch investigation.
Base model: unsloth/Qwen3-4B-Instruct-2507
Local source checkpoint:
/mnt/disks/unslothai/datta0/cache/qwen3-grpo-patch/20260605_122046_swegym_q4b-kl02-sft20k-hardmulti-ec2analog-capped-v1_ee83de6/checkpoints/best_holdout
This checkpoint is an ablation, not the promoted frontier checkpoint. It starts from the Qwen3-4B KL02 hard-multi 20k adapter and adds a capped EC2 analog SFT mixture.
Held-out SWE-Gym Moto pass@8 evaluation:
| metric | result |
|---|---|
| greedy | 9/35 |
| selected@1 | 10/35 |
| pass@8 | 11/35 |
| single-file pass@8 | 9/18 |
| multi-file pass@8 | 2/17 |
The ablation was negative against the hard-multi seed9012 frontier and should not be promoted as the best adapter.