Results
Table with columns: Correct, In language, Did not finish, J@1, J@16, J@1, released Qwen3-4B, J@1, multilingual SFT | Correct | In language | Did not finish | J@1 | J@16 | J@1, released Qwen3-4B | J@1, multilingual SFT |
|---|
| English | 68.8% | 99.8% | 6.7% | 68.8% | 88.6% | 80.4% | 71.7% |
| Spanish | 64.6% | 100.0% | 10.7% | 64.6% | 88.6% | 0.1% | 64.6% |
70 competition problems (30 AIME-2026 + 40 AMC-2023) per language from Multilingual-Math-Reasoning-Eval, 16 samples each at temperature 0.7, 24,576-token budget. A response is delivered when the answer is correct, the visible reasoning is in the requested language, and it terminates (J = C ∧ L ∧ T, scored per response). The last two columns are the released model and the six-language SFT model on the same problems.
This model was evaluated in Spanish and English only.
Quick start
from datasets import load_datasetfrom transformers import AutoModelForCausalLM, AutoTokenizer name = "96kevinli29/Qwen3-4B-SFT-Math-es-10k"tok = AutoTokenizer.from_pretrained(name)model = AutoModelForCausalLM.from_pretrained(name, torch_dtype="auto", device_map="auto") prompt = load_dataset("96kevinli29/Multilingual-Math-Reasoning-Eval", "es")["amc23"][0]["prompt"] # Spanish problem and instructiontext = tok.apply_chat_template([{"role": "user", "content": prompt}], tokenize=False, add_generation_prompt=True)inputs = tok(text, return_tensors="pt").to(model.device)out = model.generate(**inputs, max_new_tokens=24576, do_sample=True, temperature=0.7)print(tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
Qwen3 chat template with thinking on; the reasoning appears inside <think> … </think>. Loads with transformers 4.x and 5.x.
Training
Table | |
|---|
| Objective | token-level cross-entropy over the whole response, reasoning included |
| Optimizer | AdamW, lr 2e-5 with cosine decay to 0.1×, warmup ratio 0.01, weight decay 0.1 |
| Batch · context | 4 traces · up to 32,768 tokens |
| Epochs | 3 configured; epoch 1 released |
| Precision · clip | bfloat16 · gradient clip 1.0 |
| Hardware | 4 × A100-40GB on 1 node, MeluXina (LuxProvide) |
| Framework | verl SFT trainer, FSDP, full-parameter |
The other controls: English, Chinese, French, Arabic, Russian, and English at 43K.
Limitations
- A control checkpoint: one seed, one language, mathematics only, machine-translated training data.
- Like the six-language model, it can fall into repetition loops on long problems (see the Did not finish column).
Citation
@article{li2026posttraining, title = {What Does Post-Training Change in Multilingual Reasoning?}, author = {Li, Hongyang and Li, Xiao and Wu, Caesar and Danoy, Gr{\'e}goire and Bouvry, Pascal}, journal = {arXiv preprint arXiv:2609.37104}, year = {2026}} @misc{qwen3-4b-sft-math-es-10k-2026, title = {{Qwen3-4B-SFT-Math-es-10k}: Spanish-only long chain-of-thought math SFT of {Qwen3-4B}}, author = {Li, Hongyang and Li, Xiao and Wu, Caesar and Danoy, Gr{\'e}goire and Bouvry, Pascal and {Sea-Fill Community}}, year = {2026}, publisher = {Hugging Face}, howpublished = {\url{https://huggingface.co/96kevinli29/Qwen3-4B-SFT-Math-es-10k}}, note = {Maintained by Sea-Fill Community.}}
Built and released by the authors together with the Sea-Fill Community, which maintains this release.