laion/a3-rl-DCAgent_mix_h4_binary_easy-50-8B
The laion/a3-rl-DCAgent_mix_h4_binary_easy-50-8B model is an 8 billion parameter RL (SkyRL) checkpoint, fine-tuned by laion on the DCAgent/mix_h4_binary_easy task mix. It is based on the a3 GLM-SFT base model and optimized for specific reinforcement learning environments. This model demonstrates a pass@8 score of 0.672 at global_step 50, indicating its proficiency in the targeted binary easy tasks.
Loading preview...
Model Overview
laion/a3-rl-DCAgent_mix_h4_binary_easy-50-8B is an 8 billion parameter Reinforcement Learning (RL) checkpoint developed by laion. It leverages the SkyRL framework and is fine-tuned on the DCAgent/mix_h4_binary_easy task mix, building upon the a3 GLM-SFT base model. The training involved 2 epochs, concluding at global_step 67, with the checkpoint at global_step 50 selected based on the highest trailing-5 EMA of reward/avg_raw_reward.
Key Capabilities
- Reinforcement Learning Performance: Achieved an
avg_raw_rewardof 0.529 and a pass@8 score of 0.672 at the selected checkpoint (global_step 50) on theDCAgent/mix_h4_binary_easytasks. - Efficient Training: Completed training in 2 epochs, indicating focused optimization for the target environment.
- Traceability: Training-time Daytona/Harbor rollouts are available as a companion dataset at open-athena/a3-rl-DCAgent_mix_h4_binary_easy, providing insights into the policy's behavior during training.
Good For
- RL Research: Ideal for researchers and developers working on reinforcement learning problems, particularly those involving the
DCAgent/mix_h4_binary_easytask mix. - Benchmarking: Can serve as a baseline or comparison point for new RL algorithms or fine-tuning strategies within similar environments.
- Understanding RL Training: The provided training traces and logs offer valuable resources for analyzing RL policy development and performance.