TRACTRAC 価格

According to monitoring by Dongcha Beating, 'on-track sampling' in post-training of large models (i.e., training the model based on its own real-time generated data) is crucial for preventing model degradation and enhancing problem-solving abilities. Online reinforcement learning (RL) and on-track distillation (OPD) outperform traditional supervised fine-tuning (SFT) because they allow the model to optimize based on the steps it generates itself, rather than memorizing external standard answers. SFT forcibly instills standard answers, applying modification uniformly across each word, which can easily disrupt the model's original knowledge structure and lead to forgetting. In contrast, RL and OPD enable the model to find and reinforce the best steps within its own drafts. This not only avoids the cumulative error of 'writing one wrong word at the beginning and going off track later,' but also ensures that updates occur only within the areas of knowledge the model already possesses, thereby maximizing the retention of its inherent capabilities. In the 'minimal code editing' experiment, whether using SFT or RL mentors for on-track distillation, the student model's success rate (Pass@1) for writing correct code in one attempt reached 80.0% and 78.7%, respectively, both surpassing that of the mentor model. Even when the SFT mentor became significantly 'dumbed down' due to excessive fine-tuning (dropping from 0.320 to 0.286 in the LiveCodeBench coding ability test), the student model it produced still achieved a high score of 0.297, showing almost no impact from the mentor's flaws, thus proving that on-track practice can effectively filter out the mentor's bad habits. Currently, DeepSeek-V4 and GLM-5 have incorporated on-track distillation to merge expert model capabilities. In expert training, areas with clear right and wrong answers, such as coding and mathematics, are more suited for RL, while creative and knowledge-based subjective tasks are better suited for on-track distillation. Future ultimate fine-tuning algorithms will need to seek new mechanisms under the on-track training framework that combine the high efficiency of distillation (high information density) with the objectivity of RL (unbiased updates).

TRAC fluctuated by 74.03% within 24 hours, current price $0.3823

TRAC fluctuated by 52.96% within 24 hours, current price $0.462