GFT: From Imitation to Reward Fine-Tuning with Unbiased Group Advantages and Dynamic Coefficient Rectification

Create a library that integrates group-based advantage functions into standard RLHF pipelines. This allows for more stable policy updates by rectifying training coefficients based on group performance metrics.

Suggested repo: gft-trainer

"Stop training on noisy averages and start optimizing groups."

Estimated effort: 40h