Weights merged from VERL FSDP SFT checkpoint global_step_500 (500 optimizer steps, 1 epoch schedule) of Simple Self-Distillation (SSD) applied to Qwen/Qwen3-4B: sample N=20 self-generated responses from the frozen base model, then SFT on those samples.

Training data

Parquet SFT corpus (16k rows, messages column): CL-From-Nothing/RLVE-Eval20-Qwen3-4B-SSD-N20-SFT-Train.

Companion 1.7B model: CL-From-Nothing/Qwen3-1-7B-SSD-RLVE-Eval20-N20-global-step-500.

Load

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "CL-From-Nothing/Qwen3-4B-SSD-RLVE-Eval20-N20-global-step-500"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")