Qwen3-4B-OT3-law (law teacher, SFT)
Qwen3-4B-OT3-law is the law domain teacher of the MMOPD study: MMOPD/Qwen3-4B-OT3-2ep fine-tuned (supervised) on
verified long-chain-of-thought answers of a much stronger model (Qwen3.6-35B-A3B, CaseHOLD 79.7%) to legal multiple-choice
questions (CaseHOLD holding selection and bar-exam questions). Only traces whose final answer matched the gold letter were
kept (rejection sampling), at most 2 per question. It raises CaseHOLD from 63.2 to 73.9 while keeping the Qwen3 thinking
format. This is the SFT-only teacher used by the MMOPD experiments (an earlier revision of this repository held an
SFT + GRPO variant; that model reached CaseHOLD 78.3 but is not the teacher of record).
Training
| init | MMOPD/Qwen3-4B-OT3-2ep |
| data | 56,123 verified teacher traces on 28,615 prompts (CaseHOLD train 54,030 rows + bar-exam MCQ 2,093 rows), 8-gram decontaminated against the evaluation sets |
| trace teacher | Qwen3.6-35B-A3B, 8 samples per question at temperature 1.0, keep the correct ones (up to 2) |
| recipe | SFT, 2 epochs, LR 1e-5 (5% warmup), max length 20,480, flatten packing, bf16 + ZeRO-2; final checkpoint (step 104) |
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-law (this) | – | – | – | 73.9 | – | – |
| 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-law (this) | 45.4 | 42.9 | 40.8 | 37.0 | 48.2 | 29.0 |
| Qwen3-4B-OT3-2ep (student init) | 66.3 | 56.3 | 58.3 | 51.7 | 51.0 | 27.7 |
Notes
- Weights are stored in bfloat16 (the SFT checkpoint as saved). 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 teachersMMOPD/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-law"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype="auto", device_map="auto")
messages = [{"role": "user", "content": "Which of the following holdings best completes the citing text? (CaseHOLD-style question with five candidate holdings)"}]
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-law --max-model-len 40960 (the same sampling settings apply).