File size: 5,215 Bytes
fb7afac
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
---
license: other
language:
- en
task_categories:
- text-generation
tags:
- reasoning
- instruct
- math
- code
- pretraining
- front-loading
---

# MV-FrontLoad-Reasoning-Instruct

The **reasoning / instruct subset of MixtureVitae v1** that we "front-load" into pretraining
(MixtureVitae post-training / front-loading paper). 40 shards, **175.1B tokens**, 119,446,588 documents
(GPT-NeoX token counts of the data used in our runs). Text only.
Same layout as [`ontocord/MixtureVitae-v1`](https://huggingface.co/datasets/ontocord/MixtureVitae-v1): `data/<folder>/<name>.jsonl.gz`.

Each line: `{"id": ..., "metadata": ..., "text": ...}`. Use the `text` field.

## What is front-loading?
Mix this subset into a web-heavy corpus **during pretraining** (not only at the end).
Our reference blend is **42.0573 % MV-FrontLoad + 57.9427 % web** (web = Comma 0.1 in the example below).
Per-shard weights are in [`mix/blend_weights.json`](mix/blend_weights.json).

## Mix it into any web dataset
```bash
pip install datasets huggingface_hub

# Example: Comma 0.1 + MV-FrontLoad (the paper blend)
python mix/mix_frontload.py --web_dataset common-pile/comma_v0.1_training_dataset \
    --out comma_plus_mv.jsonl.gz --max_docs 1000000

# Any other web dataset: change the id (and optionally --mv_share)
python mix/mix_frontload.py --web_dataset HuggingFaceFW/fineweb-edu --web_name sample-10BT \
    --out fw_plus_mv.jsonl.gz --max_docs 1000000
```
For Megatron-LM (tokenized with GPT-NeoX; weights exactly as in our runs):
```bash
python mix/megatron_blend.py /path/to/mv_tokenized /path/to/comma_tokenized > blend.txt
```

## Files (40 shards)
| file | docs (trained) | tokens (B) |
|---|---|---|
| `data/synthetic_instruct/dart-math.jsonl.gz` | 1,426,292 | 0.60 |
| `data/synthetic_instruct/few-shot-0.jsonl.gz` | 3,577,968 | 5.97 |
| `data/synthetic_instruct/few-shot-1.jsonl.gz` | 3,482,063 | 5.88 |
| `data/synthetic_instruct/few-shot-2.jsonl.gz` | 2,845,383 | 4.83 |
| `data/formatted_text/formatted_text-0.jsonl.gz` | 998,235 | 0.59 |
| `data/formatted_text/formatted_text-1.jsonl.gz` | 723,944 | 0.43 |
| `data/synthetic_instruct/glaiveai-reasoning-0.jsonl.gz` | 2,564,555 | 4.18 |
| `data/synthetic_instruct/glaiveai-reasoning-1.jsonl.gz` | 2,556,985 | 4.18 |
| `data/synthetic_instruct/glaiveai-reasoning-2.jsonl.gz` | 2,564,005 | 4.18 |
| `data/synthetic_instruct/glaiveai-reasoning-3.jsonl.gz` | 2,565,084 | 4.18 |
| `data/synthetic_instruct/glaiveai-reasoning-4.jsonl.gz` | 2,011,319 | 5.44 |
| `data/synthetic_instruct/glaiveai-reasoning-5.jsonl.gz` | 2,558,097 | 4.18 |
| `data/synthetic_instruct/glaiveai-reasoning-6.jsonl.gz` | 2,562,654 | 4.18 |
| `data/synthetic_instruct/glaiveai-reasoning-7.jsonl.gz` | 2,563,698 | 4.18 |
| `data/synthetic_instruct/glaiveai-reasoning-8.jsonl.gz` | 2,250,828 | 3.68 |
| `data/synthetic_instruct/ling-coder-dpo-0.jsonl.gz` | 250,770 | 0.39 |
| `data/synthetic_instruct/ling-coder-sft-0.jsonl.gz` | 3,270,254 | 5.37 |
| `data/synthetic_instruct/ling-coder-sft-1.jsonl.gz` | 593,642 | 0.97 |
| `data/synthetic_instruct/ling-coder-syntheticqa-0.jsonl.gz` | 7,081,780 | 5.81 |
| `data/synthetic_instruct/ling-coder-syntheticqa-1.jsonl.gz` | 7,082,200 | 5.81 |
| `data/synthetic_instruct/ling-coder-syntheticqa-2.jsonl.gz` | 7,080,222 | 5.81 |
| `data/synthetic_instruct/ling-coder-syntheticqa-3.jsonl.gz` | 2,875,396 | 2.36 |
| `data/math/math-0.jsonl.gz` | 615,020 | 4.23 |
| `data/math/math_word_problems-0.jsonl.gz` | 587,102 | 0.46 |
| `data/synthetic_instruct/misc_instruct.jsonl.gz` | 11,057,192 | 5.51 |
| `data/synthetic_instruct/nemo_science_math-0.jsonl.gz` | 1,141,410 | 5.73 |
| `data/synthetic_instruct/nemo_science_math-1.jsonl.gz` | 25,828,968 | 14.34 |
| `data/synthetic_instruct/nemo_science_math-2.jsonl.gz` | 3,563,981 | 1.99 |
| `data/synthetic_instruct/open_thoughts-0.jsonl.gz` | 356,372 | 5.92 |
| `data/synthetic_instruct/open_thoughts-1.jsonl.gz` | 411,575 | 6.62 |
| `data/synthetic_instruct/open_thoughts-2.jsonl.gz` | 423,973 | 6.14 |
| `data/synthetic_instruct/os-q2-0.jsonl.gz` | 1,696,808 | 4.60 |
| `data/synthetic_instruct/os-q2-1.jsonl.gz` | 1,554,838 | 4.71 |
| `data/synthetic_instruct/os-q2-2.jsonl.gz` | 1,544,681 | 4.70 |
| `data/synthetic_instruct/os-q2-3.jsonl.gz` | 1,027,798 | 3.36 |
| `data/synthetic_instruct/prism_math.jsonl.gz` | 996,799 | 5.65 |
| `data/synthetic_instruct/pyedu_reasoning.jsonl.gz` | 1,609,646 | 3.10 |
| `data/synthetic_instruct/ring-lite-sft-0.jsonl.gz` | 1,411,484 | 6.78 |
| `data/synthetic_instruct/ring-lite-sft-1.jsonl.gz` | 1,535,327 | 6.75 |
| `data/synthetic_instruct/ring-lite-sft-2.jsonl.gz` | 598,240 | 1.36 |

## Provenance and caveats
- Source: MixtureVitae v1 reasoning/instruct pools, decontaminated version (`MixtureVitae-211BT-decontaminated`).
- 38 of 40 files have exactly the same number of documents as the tokenized shards we trained on.
- `formatted_text-0/1` are the **unfiltered** versions: they have more documents than the ~1.0M / 0.7M we trained on. They are <1 % of the mix.
- Decontamination is offline n-gram based and not perfect. Check against your own benchmarks.
- Each source keeps its own licence; see `metadata` and the MixtureVitae v1 card. Review before commercial use.