qwen3-8B-hanabi-step110

Qwen3-8B Hanabi RLVR Fine-tuned (Step 110)

This is a Qwen3-8B model fine-tuned using Reinforcement Learning from Verifier Rewards (RLVR) on Hanabi game tasks.

Model Details

• Base Model: Qwen/Qwen3-8B
• Training Method: RLVR (Reinforcement Learning from Verifier Rewards)
• Task: Hanabi cooperative card game
• Training Steps: 110 training steps
• Architecture: Transformer with flash attention
• Training Framework: Prime-Rel RL framework

Training Details

• Fine-tuned on Hanabi game scenarios with verifier feedback
• Uses multi-GPU setup (8x A100 GPUs)
• Optimized for strategic decision making in cooperative gameplay
• Trained with reinforcement learning from verifier rewards

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Anonymouslolol/qwen3-8B-hanabi-step110")
tokenizer = AutoTokenizer.from_pretrained("Anonymouslolol/qwen3-8B-hanabi-step110")

# Use for Hanabi game reasoning and strategic planning

Performance

• Specialized for Hanabi game strategy and reasoning
• Trained to make optimal moves in cooperative card games
• Fine-tuned for understanding game state and planning moves

Citation

Trained using Prime-Rel RL framework on Hanabi benchmark.

Files

• pytorch_model.bin: Main model weights (16GB)
• config.json: Model configuration
• tokenizer.json: Tokenizer data
• vocab.json, merges.txt: BPE tokenizer files
• generation_config.json: Generation settings
• special_tokens_map.json: Special token mappings