chord-on-policy-off-policy-harmonization

Harmonize supervised fine-tuning and reinforcement learning through dynamic weighting, balancing expert imitation and on-policy exploration to prevent response pattern disruption.

Independent GitHub snapshot · Automated classification · Original documentation governs