---
license: apache-2.0
language:
- en
task_categories:
- text-generation
tags:
- sft
- math
- code
- reasoning
- shuffled
size_categories:
- 1M<n<10M
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train_*.jsonl
---

# Mephisto-MathCode_2M

**2,000,000** non-thinking SFT examples — an even 1M/1M split of math and code
— drawn from
[openbmb/UltraData-SFT-2605](https://huggingface.co/datasets/openbmb/UltraData-SFT-2605),
filtered to English and globally shuffled.

This is a **curation** pass, not a generation one: no model produced these
answers for this dataset. All credit for the content belongs to OpenBMB. What
is added here is language filtering, an exact 1M/1M balance, and a global
shuffle so the file can be streamed without a shuffle buffer.

| `domain` | rows |
|---|---:|
| `Code` | 1,000,000 |
| `Math` | 1,000,000 |

Source config/split: `Code/no_think` and `Math/no_think`.

## Format

```json
{
  "uid": "...",
  "messages": [
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ],
  "source": "UltraData-sft-2605",
  "domain": "Code",
  "think_type": "no_think"
}
```

The source schema is preserved unchanged. Every row is exactly two turns
(`user`, `assistant`) and `think_type` is `no_think` throughout — responses are
direct answers with **no chain-of-thought block**, though math answers do show
their working as ordinary prose/LaTeX.

`domain` distinguishes the two halves, and `uid` maps back to the source row.

## Filtering

Only one filter was applied: **Chinese removal**. Rows whose prompt or answer
is more than 5% CJK characters were dropped.

| | scanned | kept | dropped (CJK) |
|---|---:|---:|---:|
| Code | 1,000,171 | 1,000,000 | 171 (0.017%) |
| Math | 1,000,154 | 1,000,000 | 154 (0.015%) |

Verified by re-running the filter over the output: 0 CJK rows and 0 duplicate
`uid`s remained. **No quality, length, or degeneracy filtering was performed** —
inspect before training.

## Shuffling

Shuffled globally across both halves with a seeded Fisher–Yates permutation
(ChaCha8, seed 42), so every shard and every prefix is representative:

| | Code | Math |
|---|---:|---:|
| shard 000 | 50.4% | 49.6% |
| shard 019 | 49.2% | 50.8% |

`take(n)` on a streaming load gives an unbiased, balanced sample without an
extra shuffle buffer.

The shuffle was done with a small memory-mapped Rust tool that permutes an
index of `(file_id, offset, length)` — 16 bytes per row, so 2M rows cost 32 MB
of RAM regardless of the 10 GB of text behind them, and line bytes go straight
from mmap to the output without ever entering the process heap. This matters
if you want to re-shuffle it yourself on a modest machine.

## Caveats

- Unfiltered for quality. The source is broadly good but nothing here has been
  checked for correctness, degeneracy, or answer length.
- Math rows are long (the 1M math rows are ~7.1 GB vs ~2.7 GB for 1M code
  rows), so a token-balanced mix is **not** 50/50 by row count.
- English-only by construction; the source split is bilingual.
- No deduplication beyond exact `uid` collisions.

## Provenance and license

All content from `openbmb/UltraData-SFT-2605` (Apache-2.0); see the source
dataset for its terms. Released under Apache-2.0.