To reproduce the SFT data pipeline (Section 3-4), I'm missing three things:
- IssueTasks data source (Table 2, rank 4) — no
generate.py in data/. How was it generated?
- Tezos 997→21K instruction rewriting prompt (Section 4) — Found
HARDEN_PROMPT in data/synthetic_harden/utils.py (gpt-4o-mini, temp=0.7), which looks like the right direction. 🙏 One clarification: this prompt focuses on "hardening" (adding constraints/complexity) — was the 997→21K expansion done by running harden_instructions() multiple times with varied seeds, or was there a separate diverse rephrasing step with a different prompt? If the former, then HARDEN_PROMPT × multiple invocations at temp=0.7 already answers it.
- Response-length upsampling formula (Section 4) — "proportionally to score" is underspecified. What's the exact formula mapping response-length to rollout count? Any cap/min/max? Total rollout budget?
To reproduce the SFT data pipeline (Section 3-4), I'm missing three things:
generate.pyindata/. How was it generated?
HARDEN_PROMPTindata/synthetic_harden/utils.py(gpt-4o-mini, temp=0.7), which looks like the right direction. 🙏 One clarification: this prompt focuses on "hardening" (adding constraints/complexity) — was the 997→21K expansion done by runningharden_instructions()multiple times with varied seeds, or was there a separate diverse rephrasing step with a different prompt? If the former, thenHARDEN_PROMPT× multiple invocations at temp=0.7 already answers it.