---
license: apache-2.0
base_model: Qwen/Qwen3-1.7B
tags:
  - biology
  - bio-posttrain
  - dna-sft
  - dna
library_name: transformers
---

# Bio-posttrain Qwen3-1.7B DNA SFT

DNA supervised fine-tuning (SFT) checkpoint from [How Post-Training Shapes Biological Reasoning Models](https://huggingface.co/collections/mims-harvard/bio-posttrain).

## Model details

- **Base model:** `Qwen/Qwen3-1.7B`
- **DNA encoder:** Evo2 `evo2_1b_base` (frozen; not included in this repo)
- **Embedding layer:** `blocks.20.mlp.l3`
- **LoRA:** rank 64, alpha 128
- **Validation loss:** 0.4687

This repo contains the **merged text LLM** (LoRA fused into base weights) plus `dna_projection.pt`.

## Files

| File | Description |
|------|-------------|
| `model.safetensors` | Merged Qwen3-1.7B weights |
| `dna_projection.pt` | Linear map from Evo2 hidden (1920) → text hidden (2048) |
| `dna_model_config.json` | DNA encoder + projection metadata |

## Loading

Use the [BioReason](https://github.com/mims-harvard/BioReason) `DNALLMModel` loader:

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

text_model = AutoModelForCausalLM.from_pretrained("mims-harvard/bio-posttrain-qwen3-1.7b-dna-sft")
tokenizer = AutoTokenizer.from_pretrained("mims-harvard/bio-posttrain-qwen3-1.7b-dna-sft")
proj = torch.load("mims-harvard/bio-posttrain-qwen3-1.7b-dna-sft/dna_projection.pt", map_location="cpu")
# Load Evo2 separately: evo2_1b_base
```

See `dna_model_config.json` for encoder settings.

## Collection

Part of the [Bio-posttrain](https://huggingface.co/collections/mims-harvard/bio-posttrain) collection.