Plans to release Stage-1 (Self-Taught SFT) code and/or checkpoints?

[FangzhiLi]

Hi, thanks for open-sourcing MTSQL-R1, really interesting work, and appreciate the Stage-2 RL code being public.

Trying to reproduce the full pipeline, I noticed Stage-1 seems incomplete here:

• No implementation of Difficulty-Aware Reject Sampling, or the LLaMA-Factory SFT config (dataset template / loss masking). run_generation_multiturn.sh also has an unresolved TODO (line 72).
• No Stage-1 checkpoint — MODEL_PATH in run_text2sql.sh points to ../llamafactory/saves/..., which isn't in this repo or on HF.
• text2sql_multiturn_w_tool.py still looks like the original GSM8k preprocessing script, not a CoSQL/SParC converter.

Any timeline for releasing the Stage-1 code, and/or a Stage-1 checkpoint?

Thanks again!