Qwen3-4B-OT3-if (instruction-following teacher, RL)

Qwen3-4B-OT3-if is the instruction-following (IF) teacher of the MMOPD study: MMOPD/Qwen3-4B-OT3-2ep trained with GRPO reinforcement learning on prompts with verifiable formatting / content constraints (IFEval-style instruction ids with programmatic checkers). It raises IFEval from 51.0 to 79.1 and IFBench from 27.7 to 58.7 while keeping math and code performance.

Training

init	MMOPD/Qwen3-4B-OT3-2ep
RL	GRPO (NeMo-RL 0.7), 200 steps, 64 prompts x 8 samples per step, LR 3e-6, no KL penalty, 32,768-token rollouts at temperature 1.0
reward	all verifiable instructions of the prompt satisfied (IFEval instruction registry checkers on the post-thinking answer)
pool	11,728 prompts from the Nemotron RL instruction-following data (verifiable-instruction slice), decontaminated against IFEval / IFBench
checkpoint	RL step 200

Evaluation

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-if (this)	70.9	–	–	58.4	58.6	–
Qwen3-4B-OT3-2ep (student init)	69.8	13.7	75.2	63.2	58.3	24.4

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-if (this)	59.2	52.5	59.2	51.4	79.1	58.7
Qwen3-4B-OT3-2ep (student init)	66.3	56.3	58.3	51.7	51.0	27.7

Notes

• Weights are stored in bfloat16 (cast from the fp32 training master weights; this is the precision in which the teacher was evaluated and served). Apache-2.0.
• Domain scores are measured at temperature 1.0 because that is the sampling regime in which these models serve as distillation teachers; general benchmarks use the Qwen3 thinking preset.
• Part of the MMOPD model family together with MMOPD/Qwen3-4B-OT3-{1ep,2ep}, MMOPD/Qwen3-1.7B-OT3-{1ep,2ep} and the other domain teachers MMOPD/Qwen3-4B-OT3-{medical,law,finance,if}.

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.

from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "MMOPD/Qwen3-4B-OT3-if"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype="auto", device_map="auto")
messages = [{"role": "user", "content": "Write a product description for a reusable water bottle in exactly three paragraphs, without using the letter z, and end with the phrase 'Stay hydrated.'"}]
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-if --max-model-len 40960 (the same sampling settings apply).