CL-From-Nothing/

RLVE-Eval20-Qwen3-4B-SSD-N20-SFT-Train

Self-generated SFT corpus for Simple Self-Distillation (SSD) with Qwen/Qwen3-4B.

800 RLVE Eval20 (filtered) prompts × 20 self-samples = 16,000 rows Sampled from frozen Qwen3-4B (vLLM, max_tokens=16384, thinking enabled). Stored as VERL MultiTurnSFTDataset parquet with a messages column.

Companion 1.7B dataset: CL-From-Nothing/RLVE-Eval20-Qwen3-1.7B-SSD-N20-SFT-Train.