STELLA-VLM-32b

STELLA-VLM-32b is a fine-tuned version of Qwen2.5-VL-32B-Instruct using Group Relative Policy Optimization (GRPO) with LoRA.

Model Details

• Base Model: Qwen/Qwen2.5-VL-32B-Instruct
• Fine-tuning Method: GRPO (Group Relative Policy Optimization) with LoRA (rank=64)
• Training Data: Scientific protocol datasets (jove_llamafactory, finebio)
• Parameters: 34B total parameters with 566M trainable LoRA parameters (1.66%)

Training Configuration

• LoRA rank: 64
• LoRA alpha: 128
• Training epochs: 3 (checkpoint saved at step 400)
• Batch size: 4
• Learning rate: 2e-4
• Reward function: Rule-based with length and repetition penalties

Usage

from transformers import AutoModelForVision2Seq, AutoProcessor
import torch

model = AutoModelForVision2Seq.from_pretrained(
    "Zaixi/STELLA-VLM-32b",
    torch_dtype=torch.bfloat16,
    trust_remote_code=True
)
processor = AutoProcessor.from_pretrained("Zaixi/STELLA-VLM-32b", trust_remote_code=True)

Fine-tuning Details

This model was fine-tuned using GRPO on scientific protocol datasets to improve instruction following and consistency in generating scientific content. The model shows improved performance on:

• Scientific protocol understanding
• Consistent response generation
• Following detailed instructions
• Multimodal reasoning tasks

License

Apache 2.0