Qwen3-4B-OT3-finance (finance teacher, RL)
Qwen3-4B-OT3-finance is the finance domain teacher of the MMOPD study: MMOPD/Qwen3-4B-OT3-2ep trained with GRPO
reinforcement learning on financial-report question answering with numeric answers (FinQA and TAT-QA training questions),
using a verifiable numeric-match reward. It raises FinQA from 58.3 to 74.6.
Training
| init | MMOPD/Qwen3-4B-OT3-2ep |
| RL | GRPO (NeMo-RL 0.7), 200 steps, 32 prompts x 8 samples per step, LR 3e-6, no KL penalty, 16,384-token rollouts at temperature 1.0 |
| reward | numeric answer match (DocMath-Eval comparator: 0.15% relative tolerance, percent / scale normalisation) |
| pool | 7,565 prompts: TAT-QA train 4,506 + FinQA train 3,059 (numeric-answer questions over report tables and text), decontaminated against the evaluation sets |
| 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-finance (this) | 69.1 | – | – | 60.8 | 74.6 | – |
| 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-finance (this) | 50.8 | 55.8 | 52.1 | 47.8 | 34.6 | 24.3 |
| 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 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-finance"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype="auto", device_map="auto")
messages = [{"role": "user", "content": "Given the income statement excerpt below, what was the percentage change in operating income from 2019 to 2020? ..."}]
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-finance --max-model-len 40960 (the same sampling settings apply).