Clarifications on Roadmap

[zyzhang1130]

Hi, I am very interested in this project and would like to contribute, but before that I have some clarifications want to make regarding the roadmap:

1. In the repo's main page README Plan of attack section Step 2 and 3, I notice the reproduction pipeline looks different from the Deepseek R1 report. In particular, there is no mentioning of 'RL reasoning data' in the report, and I don't understand of the purpose of it since the RL part is supposed to be data-free (i.e., solely rely on a rule-based verifier to assign reward for the outcome).
2. At Step 3: R1 was training with 2 rounds of RL and a SFT in between, why here there is only one round of RL after the SFT?

Appreciate if some main contributor can answer my questions. Also I would like to ask if there is any more effective way to communicate (e.g., Discord) available? Thanks.