Supervised fine-tune (SFT) of Qwen/Qwen3-4B-Base on allenai/Dolci-Instruct-SFT — the ~2.15M-example instruction mixture used for the Olmo-3 Instruct SFT stage. Uses the standard Qwen3 chat template; compatible with transformers Qwen3ForCausalLM.

Training


Base model	Qwen/Qwen3-4B-Base (~4.0B params)
SFT data	allenai/Dolci-Instruct-SFT (~2.15M examples)
Epochs	~2.03 (34,000 steps)
Global batch size	128
Max sequence length	4096
Optimizer	AdamW (weight_decay=0.0), grad-norm clip 1.0
LR schedule	peak 2e-6, cosine decay to 0, 500 warmup steps
Loss	assistant-only (prompt/system/tool tokens masked)
Hardware	v6e-64 (FSDP=64) TPU

Evaluation

Scores at step 34,000 with enable_thinking=False. IFEval / IFBench report strict instruction-level accuracy; GSM8K / MATH-500 report accuracy; Multi-IF is the mean of turn_{1,2,3} strict instruction-level; MultiChallenge is the overall score (GPT-4o judge).

Model	IFEval	IFBench	GSM8K	MATH-500	Multi-IF	MultiChallenge
Qwen3-4B-SFT-Dolci (this model)	83.81	20.64	89.69	68.80	61.78	16.06
Qwen3-8B-SFT-Dolci	84.29	22.67	91.36	73.00	61.71	15.41
Qwen3-14B-SFT-Dolci	84.17	26.45	94.24	74.40	66.61	15.30

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

repo = "yxli2123/Qwen3-4B-SFT-Dolci"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype=torch.bfloat16, device_map="auto")

messages = [{"role": "user", "content": "Give me three tips for writing clear code."}]
ids = tok.apply_chat_template(messages, add_generation_prompt=True,
                              enable_thinking=False, return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=256)
print(tok.decode(out[0][ids.shape[-1]:], skip_special_tokens=True))