s-angheben/SpeciaRL_qwen2_5vl-7b_rft

SpeciaRL_rft

This repository provides the LoRA adapter for the rft fine-tuned model introduced in SpeciaRL. Built on top of Qwen/Qwen2.5-VL-7B-Instruct.

Training hyperparameters

• —algorithm: GRPO
• —learning_rate: 3e-5
• —trainbatchsize: 256
• —maxpromptlength: 2048
• —maxresponselength: 2048
• —lora_rank: 64
• —lora_alpha: 32
• —target_modules: all-linear (visual layers excluded)
• —klloss: True (coef=0.01, type=lowvar_kl)
• —rollout_n: 10
• —num_gpus: 4
• —total_epochs: 15

Framework versions

• —VERL
• —PEFT 0.17.1
• —Transformers 4.57.0
• —PyTorch 2.6.0+cu124