SpeciaRL
Collection
Collection of Specificity-aware reinforcement learning for fine-grained open-world classification • 5 items • Updated
SpeciaRL_SpeciaRL_mixed
This repository provides the LoRA adapter for SpeciaRL trained using the mixture of all available domains, more information in SpeciaRL. Built on top of Qwen/Qwen2.5-VL-7B-Instruct.
Training hyperparameters
- algorithm: GRPO
- reward_function: best prediction (dedup)
- learning_rate: 3e-5
- train_batch_size: 256
- max_prompt_length: 2048
- max_response_length: 2048
- lora_rank: 64
- lora_alpha: 32
- target_modules: all-linear (visual layers excluded)
- kl_loss: True (coef=0.01, type=low_var_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
- Downloads last month
- 2
Model tree for s-angheben/SpeciaRL_qwen2_5vl-7b_SpeciaRL_mixed
Base model
Qwen/Qwen2.5-VL-7B-Instruct