math-code-mixtral-8x7b

Math & Code Mixtral-8x7B

A LoRA fine-tune of Mixtral-8x7B-Instruct-v0.1 for mathematical word problems and short code generation, built with Adaption Labs' AutoScientist for the AutoScientist Challenge (Math & Code category).

Results

Head-to-head win rate against the base model:

These are wins in a paired comparison, not accuracy percentages.

An earlier checkpoint trained on 2,050 rows without augmentation scored 56–44 on the category evaluation. Expanding the training set with domain and general-purpose augmentation, and raising LoRA capacity to rank 64, gained 4 points. The platform's guidance suggested a larger gain from crossing 20,000 datapoints; the observed improvement was smaller than that.

Usage

The archive unpacks flat, so give it its own directory:

mkdir -p math-adapter
tar --zstd -xf math-code-mixtral-8x7b-weights.tgz -C math-adapter

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE = "mistralai/Mixtral-8x7B-Instruct-v0.1"
ADAPTER = "./math-adapter"

tokenizer = AutoTokenizer.from_pretrained(ADAPTER)
model = AutoModelForCausalLM.from_pretrained(
    BASE, torch_dtype=torch.bfloat16, device_map="auto"
)
model = PeftModel.from_pretrained(model, ADAPTER)
model.eval()

messages = [{"role": "user", "content":
    "Jean eats one donut per 2 pages she writes. If she writes 12 pages and each "
    "donut has 150 calories, how many calories does she eat?"}]
inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=256, do_sample=False)
print(tokenizer.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))

Hardware: Mixtral-8x7B is a 46.7B-parameter mixture-of-experts model — roughly 94 GB in bf16, or about 25 GB in 4-bit. The adapter is 208 MB.

Output style: trained to show brief working then state the answer explicitly, ending with "The answer is X."

Training

Training curve

Training loss fell from 1.248 to 0.764. Evaluation loss was still declining at epoch 5, so this run did not overfit — the last epoch still gained, if marginally.

Dataset

flamiinngo/math-code-qa — 5,200 rows (3,600 math, 1,600 code), every math answer ending in a result verified against the upstream expected_answer column. Derived from nvidia/OpenMathInstruct-2 and sahil2801/CodeAlpaca-20k, both CC-BY-4.0. The training run additionally used AutoScientist domain and general-purpose augmentation on top of this base.

Limitations

• It can produce confident wrong arithmetic. Fine-tuning improved the base model's working, but this is not a calculator. Check any result that matters.
• Scope is school and early-undergraduate level — word problems, algebra, combinatorics, geometry. Not competition mathematics.
• Code output is untested. The code training data was filtered for length, not executed. Treat generated code as a draft.
• Win rate is not accuracy. It measures preference against one base model on one evaluation, not correctness in absolute terms.
• English only.

License

The adapter is Apache 2.0, matching the base model. The training data is CC-BY-4.0 and requires attribution — see the dataset card.

Acknowledgements

• Adaption Labs — AutoScientist platform and challenge
• NVIDIA and sahil2801 — upstream open datasets
• Mistral AI — Mixtral-8x7B base model