---
license: apache-2.0
language:
- en
task_categories:
- text-generation
tags:
- math
- arithmetic
- synthetic
- pretraining
- small-language-models
size_categories:
- 100M<n<1B
configs:
- config_name: math-story
  data_files:
  - split: train
    path: data/math-story/*.parquet
- config_name: math-rel
  data_files:
  - split: train
    path: data/math-rel/*.parquet
- config_name: math-rel-nf
  data_files:
  - split: train
    path: data/math-rel-nf/*.parquet
- config_name: math-am-nf
  data_files:
  - split: train
    path: data/math-am-nf/*.parquet
- config_name: math-am2-nf
  data_files:
  - split: train
    path: data/math-am2-nf/*.parquet
- config_name: math-am3
  data_files:
  - split: train
    path: data/math-am3/*.parquet
- config_name: math-am3-nf
  data_files:
  - split: train
    path: data/math-am3-nf/*.parquet
- config_name: math-frames
  data_files:
  - split: train
    path: data/math-frames/*.parquet
- config_name: math-frames2
  data_files:
  - split: train
    path: data/math-frames2/*.parquet
- config_name: math-frames2w
  data_files:
  - split: train
    path: data/math-frames2w/*.parquet
- config_name: math-frames3
  data_files:
  - split: train
    path: data/math-frames3/*.parquet
- config_name: math-topics
  data_files:
  - split: train
    path: data/math-topics/*.parquet
- config_name: math-topics-frames
  data_files:
  - split: train
    path: data/math-topics-frames/*.parquet
- config_name: math-expr
  data_files:
  - split: train
    path: data/math-expr/*.parquet
- config_name: math-words
  data_files:
  - split: train
    path: data/math-words/*.parquet
- config_name: math-formats
  data_files:
  - split: train
    path: data/math-formats/*.parquet
- config_name: math-ws
  data_files:
  - split: train
    path: data/math-ws/*.parquet
- config_name: math-neg
  data_files:
  - split: train
    path: data/math-neg/*.parquet
- config_name: math-v8
  data_files:
  - split: train
    path: data/math-v8/*.parquet
---

# TinyBrainBot 3M math data

Every arithmetic dataset used to train the TinyBrainBot 3M math models (the 3M-Base punch runs and
[3M-Loop](https://huggingface.co/nkthebass/tinybrainbot-3m-loop) series). "Used by" names the public `nkthebass/tinybrainbot-*` models. Each set is its own folder / config:

```python
from datasets import load_dataset
ds = load_dataset("nkthebass/tinybrainbot-3m-math-data", "math-expr", split="train")
```

Every number and answer is computed in Python. Where a set uses LLM-written sentence frames, the LLM only wrote the
wording with placeholders; the numbers were filled in and solved by Python. No benchmark train or test data is included.

| Config | What it is | How it was made | Used by | Text | Rows |
|---|---|---|---|--:|--:|
| `math-story` | Story arithmetic in many sentence shapes: add/subtract, groups, sharing, money, time, measurement, area/perimeter, fractions, patterns, tallies | tiny3m/gen_math_story.py (Python templates) | 3m-base, 3m-loop, 3m-loop-v2, 3m-loop-v3, 3m-loop-v4 | 6000M | 13,963,297 |
| `math-rel` | Relational arithmetic: 'X is 30 cm taller than Y', older/younger, more/fewer, with the result stated | Python templates | 3m-loop, 3m-loop-v2, 3m-loop-v3, 3m-loop-v4 | 1200M | 1,643,821 |
| `math-rel-nf` | math-rel without the worked numbers in brackets | Python templates | 3m-loop, 3m-loop-v2, 3m-loop-v3 | 1200M | 1,641,096 |
| `math-am-nf` | ArithMark-1-style word problems (built to resemble ArithMark phrasing - see notes) | Python templates | 3m-base (punch), early 3M-Loop runs | 1200M | 2,432,712 |
| `math-am2-nf` | ArithMark-2-style problems (built to resemble ArithMark phrasing - see notes) | Python templates | early 3M-Loop runs | 1200M | 2,407,200 |
| `math-am3` | ArithMark-3-style continuations (built to resemble ArithMark phrasing - see notes) | Python templates | early 3M-Loop runs | 1200M | 2,440,467 |
| `math-am3-nf` | math-am3 without worked steps | Python templates | early 3M-Loop runs | 1200M | 2,404,970 |
| `math-frames` | Story problems in LLM-written sentence frames; every number filled in and computed by Python | tiny3m/gen_math_frames.py, frames by Ling-3.0-Flash | an early 3M-Loop run | 2640M | 2,524,390 |
| `math-frames2` | Same idea, two writer families | frames by Ling-3.0-Flash + Mistral-Small-3.2 | 3m-loop | 3795M | 3,772,927 |
| `math-frames2w` | math-frames2 with numbers written as words | frames by Ling-3.0-Flash + Mistral-Small-3.2 | 3m-loop-v2, 3m-loop-v3, 3m-loop-v4 | 1650M | 1,566,265 |
| `math-frames3` | More frames, plus optional irrelevant-number distractors | frames by Ling-3.0-Flash + Mistral Nemo | 3m-loop-v4 | 1980M | 1,825,387 |
| `math-topics` | 23 grade-school topics ArithMark-3 lacks: remainders, unit conversion, elapsed time, days of the week, rounding, place value, fractions, percent, averages, 3-step, dozen, odd/even, multiples, money in words | tiny3m/gen_math_topics.py (Python templates) | 3m-loop-v4 | 690M | 1,162,281 |
| `math-topics-frames` | The topics above in LLM-written frames | tiny3m/gen_math_frames.py --skills topics | 3m-loop-v4 | 1155M | 1,181,300 |
| `math-expr` | Bare expressions: order of operations, parentheses, negatives, 2-5 operands, ~19 forms | tiny3m/gen_math_expr.py | 3m-loop-v2, 3m-loop-v3, 3m-loop-v4 | 1560M | 3,381,454 |
| `math-words` | Numbers written as words, both directions, inside sums and comparisons | tiny3m/gen_math_words.py | 3m-loop-v2, 3m-loop-v3, 3m-loop-v4 | 680M | 1,066,765 |
| `math-formats` | Large numbers, thousands commas, money, decimals, mixed into short stories | tiny3m/gen_math_formats.py | 3m-loop-v3, 3m-loop-v4 | 1920M | 1,682,618 |
| `math-ws` | Whitespace variants before numbers (double spaces, tabs, newlines) | tiny3m/gen_math_ws.py | 3m-loop-v4 | 1200M | 1,580,637 |
| `math-neg` | Negative numbers: 8 sign shapes, temperatures, comparisons | tiny3m/gen_math_neg.py | 3m-loop-v4 | 300M | 501,552 |
| `math-v8` | Unknown positions (? + 5 = 12), irrelevant-detail distractors, layouts, sequences, small tables, comparisons | tiny3m/gen_math_v8.py | 3m-loop-v4 | 1610M | 2,180,515 |

## Notes from training with it

- **`math-am*` were built to resemble ArithMark phrasing.** They raise ArithMark scores much more than arithmetic in other
  wordings (a 3M on one phrasing family solved 99% in that phrasing and 10-30% elsewhere). They were dropped after
  the early 3M-Loop runs.
- **Story-style sets (`math-story`, `math-frames*`) lift ArithMark-3 about 3x faster than held-out arithmetic.** Our later
  general models use only a pinch of them.
- **`math-topics` / `math-topics-frames`** cover skills ArithMark-3 does not test. Training on them also inflates any
  test built from the same generator, so evaluate with held-out wordings.
- `math-ws` exists because lm-eval's default target delimiter adds a second space before the answer; models that never
  saw that collapsed on it.

## License

Apache-2.0. Frames were written by Ling-3.0-Flash (inclusionAI), Mistral-Small-3.2 and Mistral Nemo (Mistral AI); all
numbers and answers were generated by our own Python code.