An analysis of DeepSeek's R1-Zero and R1

Re:An analysis of DeepSeek's R1-Zero and R1 by spyckie2 on Wednesday January 29, 2025 @08:49PM (#42873838)

Yes, but I meant it slightly differently than the distills.

The idea is to create the next gen SOTA non reasoning model with synthetic reasoning training data.

[ Parent | on HN ]

Re:An analysis of DeepSeek's R1-Zero and R1 by entropicdrifter on Thursday January 30, 2025 @12:23PM (#42879898)

So you mean something like, "what if the baseline, off-the-cuff response for the next-gen models was tuned based on the results of the reasoning model excluding the reasoning itself?"

[ Parent | on HN ]

Re:An analysis of DeepSeek's R1-Zero and R1 by spyckie2 on Thursday January 30, 2025 @04:17PM (#42882278)

Exactly, albeit it may need the reasoning later to form the proper foundational logic in the weights.

[ Parent | on HN ]