mims-harvard/bio-posttrain-qwen3-1.7b-dna-rl
The mims-harvard/bio-posttrain-qwen3-1.7b-dna-rl model is a 1.7 billion parameter Qwen3-based language model developed by mims-harvard, specifically fine-tuned using DNA reinforcement learning (GRPO). This model is part of a collection exploring how post-training methods shape biological reasoning models. It integrates a frozen Evo2 DNA encoder and is optimized for tasks requiring biological reasoning, particularly those involving DNA sequence understanding.
Loading preview...
Model Overview
The mims-harvard/bio-posttrain-qwen3-1.7b-dna-rl is a specialized language model built upon the Qwen/Qwen3-1.7B base architecture. Developed by mims-harvard, this model has undergone DNA reinforcement learning (GRPO) as part of a research initiative to understand how post-training methodologies influence biological reasoning capabilities in LLMs. It is a key component of the Bio-posttrain collection.
Key Characteristics
- Base Model: Utilizes
Qwen/Qwen3-1.7Bas its foundational language model. - DNA Integration: Incorporates a frozen
evo2_1b_baseDNA encoder, with embeddings extracted fromblocks.20.mlp.l3. - Training Methodology: Fine-tuned using a combination of SFT (sft256 / LoRA rank 16) and GRPO (rank 16, alpha 32) at checkpoint step 1156.
- Components: The repository provides the merged text LLM along with
dna_projection.ptfor DNA-related functionalities.
Use Cases
This model is particularly suited for research and applications focused on:
- Biological Reasoning: Tasks that require an understanding of biological sequences, especially DNA.
- Genomic Analysis: Potential applications in interpreting and generating insights from DNA data.
- Biomedical Research: Exploring the intersection of language models and biological sciences through its specialized training.