Downloads · 30 days
19
21% of all-time downloads
burtenshaw/qwen3-0.6b-mbpp-sft-lr1e4-r16-best-len2k
qwen3-0.6b-mbpp-sft-lr1e4-r16-best-len2k is a text generation model from burtenshaw. Use it when you need the model to write or continue text. It is set up for peft.
This is a PEFT LoRA adapter for Qwen/Qwen3-0.6B trained with TRL SFT on the sanitized MBPP train split. The prompt/completion format mirrors inspectevals/mbpp: three MBPP few-shot examples, task prompt, tests, then a…
Downloads · 30 days
19
21% of all-time downloads
All-time downloads
90
Public
Repo size
153 MB
Likes
0
Public
Click a slice to open those files.
.safetensors40.4 MB · 78%
From the Hugging Face model README
This is a PEFT LoRA adapter for Qwen/Qwen3-0.6B trained with TRL SFT on the sanitized MBPP train split. The prompt/completion format mirrors inspect_evals/mbpp: three MBPP few-shot examples, task prompt, tests, then a Python completion.
Same Inspect MBPP evaluation command, seed, sampling, and JSON extraction were used for base and adapter.
| Model | Inspect MBPP extracted accuracy | stderr | Eval Job |
|---|---|---|---|
Qwen/Qwen3-0.6B | 0.2607003891050584 | 0.027438546123673805 | 6a3564cc3093dba73ce2a4d1 |
Qwen/Qwen3-0.6B + this adapter | 0.3852140077821012 | 0.030415365154196253 | 6a3567c93093dba73ce2a55d |
Absolute gain: +0.1245136186770428.
Note: the temporary evaluation summary script collapsed duplicate Inspect metric names and retained the final MBPP reducer value. The base and adapter were compared with the same extraction protocol.
SFTTrainer with completion-only loss and PEFT LoRA.Qwen/Qwen3-0.6B at commit c1899de289a04d12100db370d81485cdf75e47ca.google-research-datasets/mbpp, sanitized config, train split, revision 4bb6404fdc6cacfda99d4ac4205087b89d32030c.eval_loss (0.7686 at step 25).6a3564dd953ed90bfb944aca.burtenshaw/mbpp-sft-trackio.This adapter is a small MBPP format-recovery experiment, not a general coding model. It should be evaluated on any downstream coding benchmark before use.