GENTEL-Lab/Fold2Reason-Qwen3.5-9B

Fold2Reason Qwen3.5-9B

The three main Fold2Reason seed checkpoints are packaged together here. Each seed has a PEFT LoRA adapter and its matching workspace. The frozen geometry decoder is shared by all three seeds. These are the paper's full FoldingCorpus plus geometry results at epoch 3 / step 375, trained on 1,000 proteins. The legacy experiment name in the source manifest is m3_no_retrieval.

Path	Contents
seed-20260729/	Adapter and matching workspace
seed-20260803/	Adapter and matching workspace
seed-20260804/	Adapter and matching workspace
decoder/	Shared frozen coordinate and distogram heads

For language-only inference, use the selected seed's adapter/ directory with the official Qwen/Qwen3.5-9B base model. The base model weights and tokenizer are not included here. The paper uses base revision c202236235762e1c871ad0ccb60c8ee5ba337b9a.

from pathlib import Path
from huggingface_hub import snapshot_download
from peft import PeftModel
from transformers import AutoModelForCausalLM

repo = Path(snapshot_download(
    "GENTEL-Lab/Fold2Reason-Qwen3.5-9B",
    allow_patterns=["seed-20260729/*"],
))
base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3.5-9B", revision="c202236235762e1c871ad0ccb60c8ee5ba337b9a"
)
model = PeftModel.from_pretrained(base, repo / "seed-20260729" / "adapter")

For structural readout, download decoder/ as well and pair the chosen seed's workspace with its adapter. The evaluation guide gives the full procedure and required FoldBench cache. Do not interchange workspaces between seeds. checkpoint_manifest.json records the fixed endpoint and source file hashes. SOURCE_PROVENANCE.json records SHA-256 hashes of the files in this repository's source archive.

Paper: https://huggingface.co/papers/2609.38879
Code: https://github.com/GENTEL-lab/Fold2Reason
Source: the public project dataset.zip, SHA-256 97ef38130fec0b7caaa323022d5b7bef620e7b26dfa42a1da4713066806dd90b.

LICENSE-Qwen.txt is the upstream Qwen license notice. It does not assign a license to Fold2Reason-produced adapter, workspace, or decoder files.