README.md

---
license: apache-2.0
task_categories:
  - text-generation
language:
  - en
tags:
  - math
  - reasoning
  - mimo
  - math500
  - MATH
  - vllm
  - sft
size_categories:
  - 10K<n<100K
---

# MiMo SFT2 — MATH-500 responses (128× sampling, temp 0.6)

Model responses generated with an **SFT2 MiMo-7B** model
([`tequila3009/sft2_mimo`](https://huggingface.co/tequila3009/sft2_mimo), weights under the `sft2_mimo/` subdir)
on the **MATH-500** problem set.

## Generation setup

| | |
|---|---|
| Model | SFT2 MiMo-7B — [`tequila3009/sft2_mimo`](https://huggingface.co/tequila3009/sft2_mimo) (`sft2_mimo/`) |
| Dataset | MATH-500 — 500 problems |
| Samples per problem | **128** |
| Total responses | **64,000** |
| Temperature | 0.6 |
| top_p | 0.95 |
| top_k | -1 (disabled) |
| max_tokens | 16384 |
| Engine | vLLM, TP=8 on 8× A100 |

Each problem is wrapped in a natural-language chain-of-thought prompt that asks
the model to reason step by step and put the final answer in `\boxed{...}`.

## File

- `sft2_mimo_math500_responses_t0.6_20260707_134307.json` (~657 MB)

## Structure

```json
{
  "model": "sft2_mimo",
  "dataset": "math500.jsonl",
  "sampling_params": {"temperature": 0.6, "top_p": 0.95, "top_k": -1, "max_tokens": 16384},
  "total_questions": 500,
  "responses_per_question": 128,
  "generation_time": 34920.0,
  "results": [
    {"question": "<problem text>", "responses": ["<sample 1>", "... 128 samples ..."]}
  ]
}
```

`results` is aligned to the MATH-500 order; `results[i]["responses"]` holds the
128 samples for problem `i`.

## Notes

- All 64,000 responses are non-empty (0 empty completions).
- Responses are raw model text (reasoning + `\boxed{}` answer), not verified for correctness.

## Loading

```python
import json
data = json.load(open("sft2_mimo_math500_responses_t0.6_20260707_134307.json"))
print(data["total_questions"], data["responses_per_question"])   # 500 128
q0 = data["results"][0]
print(q0["question"])
print(q0["responses"][0])
```