Adapter details
Table with columns: Setting, Value| Setting | Value |
|---|
| Selected checkpoint | 2 completed epochs; 76 optimizer steps |
| LoRA rank / alpha / dropout | 16 / 32 / 0 |
| Adapter parameters | 116,727,808 |
| Adapted modules | 496 linear modules across all 64 text decoder layers |
| Precision | BF16 base; FP32 adapter checkpoint |
Adapters cover the text model's full-attention projections, linear-attention
projections, and MLP gate/up/down projections.
Tested software versions: Transformers 5.19.0 and PEFT 0.19.1.
import torchfrom peft import PeftModelfrom transformers import AutoTokenizer, Qwen3_5ForConditionalGeneration base_id = "Qwen/Qwen3.8-27B"base_revision = "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0" tokenizer = AutoTokenizer.from_pretrained(base_id, revision=base_revision)base = Qwen3_5ForConditionalGeneration.from_pretrained( base_id, revision=base_revision, dtype=torch.bfloat16, device_map="auto",)model = PeftModel.from_pretrained( base, "ykharkov/Qwen3.8-27B-LeanZip", is_trainable=False).eval()
The base uses the Qwen3.5-family architecture. Tokenizer and chat-template assets
come from the pinned base model.
Provide the existing Lean theorem and proof together with a request for a
shorter, equivalent proof body.
Evaluation and provenance
No benchmark score is reported for this selected checkpoint. The training data
includes benchmark tasks, so those tasks are not an independent held-out
evaluation set.
training_metadata.json records the selected checkpoint and adapter properties.
The adapter weight SHA-256 is:
022bdb626fd2d5d6e8f8e77a3c6db29f6c7a167bb85acddb6e26fdbb4177dcfb