Modelpublic

MMOPD/Qwen3-4B-OT3-2ep

sourceHugging Faceapache-2.0updated 13d agoView on Hugging Face

0likes203downloads

Model Card

Qwen3-4B-OT3 (2 epochs)

Qwen3-4B-OT3-2ep is Qwen3-4B-Base fine-tuned on OpenThoughts3-1.2M (the full 1.2M-example long-chain-of-thought SFT set: math, code and science reasoning traces) — an open re-creation of the OpenThinker3 recipe at the 4B scale. It is a thinking model: every answer starts with a <think> block. The checkpoint is a general reasoning student that the MMOPD project uses as the starting point for domain teachers and for on-policy distillation experiments.

This repository holds the final checkpoint after 2 epochs (step 4,390). The 1-epoch checkpoint is MMOPD/Qwen3-4B-OT3-1ep.

Training


base model	Qwen3-4B-Base
data	OpenThoughts3-1.2M (`open-thoughts/OpenThoughts3-1.2M`), all 1.2M rows, Qwen3 chat template with thinking
sequence length	16,384 tokens, sequence packing (flatten, no cross-example attention)
epochs	2 (4,390 optimizer steps in total, 2,195 per epoch)
optimizer	AdamW, peak LR 8e-5, 5% warmup, global batch 512 packed sequences (about 8.1M tokens per step)
precision	bf16 compute, ZeRO-2 data parallel (transformers 4.57 / trl 0.29 / DeepSpeed) on A100-80GB
final train loss	0.868

Evaluation

General benchmarks (Qwen3 thinking preset: temperature 0.6, top-p 0.95, top-k 20; 32,768 max new tokens; AIME = avg@8, LiveCodeBench v6 / IFEval / IFBench = 1 sample; scores in %):

Model	AIME24	AIME25	AIME26	LiveCodeBench v6	IFEval	IFBench
Qwen3-4B-OT3-2ep (this)	66.3	56.3	58.3	51.7	51.0	27.7
Qwen3-4B-OT3-1ep	60.4	51.3	55.4	47.1	46.8	27.0

Domain benchmarks (temperature 1.0, top-p 1.0, long generation budget; accuracy in %):

Model	MedQA	MedXpertQA	PubMedQA	CaseHOLD	FinQA	TAT-QA (EM)
Qwen3-4B-OT3-2ep (this)	69.8	13.7	75.2	63.2	58.3	24.4

Notes

—Apache-2.0, like the Qwen3 base models and OpenThoughts3.
—Part of the MMOPD (multi-teacher on-policy distillation) model family: the domain teachers MMOPD/Qwen3-4B-OT3-{medical,law,finance,if} start from MMOPD/Qwen3-4B-OT3-2ep.

How to use

The models keep the Qwen3 chat template and thinking format (<think> ... </think> before the answer). Use enable_thinking=True and sampling (not greedy); the evaluations below used a 32k-token generation budget.

python

from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "MMOPD/Qwen3-4B-OT3-2ep"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype="auto", device_map="auto")
messages = [{"role": "user", "content": "How many positive integers n < 1000 have the property that n^2 + 1 is divisible by 5?"}]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=True)
out = model.generate(**tok(text, return_tensors="pt").to(model.device), max_new_tokens=32768,
                     do_sample=True, temperature=0.6, top_p=0.95, top_k=20)
print(tok.decode(out[0], skip_special_tokens=True))

vLLM: vllm serve MMOPD/Qwen3-4B-OT3-2ep --max-model-len 40960 (the same sampling settings apply).