tinymix: 100M, 10M, 1M
Browse files- README.md +221 -0
- code/assemble.py +87 -0
- code/baselines.py +71 -0
- code/build.py +378 -0
- code/mix.py +72 -0
- code/publish.py +257 -0
- code/requirements.txt +10 -0
- code/sources.py +193 -0
- code/train.py +180 -0
- data/100M/train.parquet +3 -0
- data/100M/validation_books.parquet +3 -0
- data/100M/validation_dialogue.parquet +3 -0
- data/100M/validation_expository.parquet +3 -0
- data/100M/validation_narrative.parquet +3 -0
- data/100M/validation_news.parquet +3 -0
- data/100M/validation_poetry.parquet +3 -0
- data/10M/train.parquet +3 -0
- data/10M/validation_dialogue.parquet +3 -0
- data/10M/validation_expository.parquet +3 -0
- data/10M/validation_narrative.parquet +3 -0
- data/1M/train.parquet +3 -0
- data/1M/validation_dialogue.parquet +3 -0
- data/1M/validation_expository.parquet +3 -0
- data/1M/validation_narrative.parquet +3 -0
README.md
ADDED
|
@@ -0,0 +1,221 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- en
|
| 4 |
+
license: other
|
| 5 |
+
license_name: mixed-per-source
|
| 6 |
+
license_link: https://huggingface.co/datasets/jncarlo/tinymix#sources-and-licensing
|
| 7 |
+
pretty_name: tinymix
|
| 8 |
+
size_categories:
|
| 9 |
+
- 100K<n<1M
|
| 10 |
+
task_categories:
|
| 11 |
+
- text-generation
|
| 12 |
+
tags:
|
| 13 |
+
- small-language-models
|
| 14 |
+
- synthetic-data
|
| 15 |
+
- curriculum
|
| 16 |
+
- tinystories
|
| 17 |
+
- babylm
|
| 18 |
+
configs:
|
| 19 |
+
- config_name: 1M
|
| 20 |
+
data_files:
|
| 21 |
+
- split: train
|
| 22 |
+
path: data/1M/train.parquet
|
| 23 |
+
- split: validation_dialogue
|
| 24 |
+
path: data/1M/validation_dialogue.parquet
|
| 25 |
+
- split: validation_expository
|
| 26 |
+
path: data/1M/validation_expository.parquet
|
| 27 |
+
- split: validation_narrative
|
| 28 |
+
path: data/1M/validation_narrative.parquet
|
| 29 |
+
- config_name: 10M
|
| 30 |
+
data_files:
|
| 31 |
+
- split: train
|
| 32 |
+
path: data/10M/train.parquet
|
| 33 |
+
- split: validation_dialogue
|
| 34 |
+
path: data/10M/validation_dialogue.parquet
|
| 35 |
+
- split: validation_expository
|
| 36 |
+
path: data/10M/validation_expository.parquet
|
| 37 |
+
- split: validation_narrative
|
| 38 |
+
path: data/10M/validation_narrative.parquet
|
| 39 |
+
- config_name: 100M
|
| 40 |
+
data_files:
|
| 41 |
+
- split: train
|
| 42 |
+
path: data/100M/train.parquet
|
| 43 |
+
- split: validation_books
|
| 44 |
+
path: data/100M/validation_books.parquet
|
| 45 |
+
- split: validation_dialogue
|
| 46 |
+
path: data/100M/validation_dialogue.parquet
|
| 47 |
+
- split: validation_expository
|
| 48 |
+
path: data/100M/validation_expository.parquet
|
| 49 |
+
- split: validation_narrative
|
| 50 |
+
path: data/100M/validation_narrative.parquet
|
| 51 |
+
- split: validation_news
|
| 52 |
+
path: data/100M/validation_news.parquet
|
| 53 |
+
- split: validation_poetry
|
| 54 |
+
path: data/100M/validation_poetry.parquet
|
| 55 |
+
---
|
| 56 |
+
|
| 57 |
+
# tinymix
|
| 58 |
+
|
| 59 |
+
Small, **genre-balanced English pretraining corpora** in three sizes (about 1M, 10M and 100M words), built for training
|
| 60 |
+
very small language models. The mix combines synthetic child-level stories and dialogue, simplified real text,
|
| 61 |
+
encyclopedic and didactic text, books, poetry and news, all filtered toward clear, self-contained documents. It is
|
| 62 |
+
inspired by TinyStories, SimpleStories, TinyHelen and the BabyLM challenge.
|
| 63 |
+
|
| 64 |
+
What is different from using those datasets directly:
|
| 65 |
+
|
| 66 |
+
- **Several kinds of text in one mix** (narrative, dialogue and chat, expository, books, poetry, news), not only stories or only conversations.
|
| 67 |
+
- **Three sizes built with the same recipe**, with complexity limits that grow with size.
|
| 68 |
+
- **No non-commercial sources.** TinyChat (CC-BY-NC) and the BabyLM corpus (no clear license, includes subtitle and child-speech data) were left out on purpose.
|
| 69 |
+
- **Provenance per document** (`source`, `genre`) and **validation splits per genre**, to see which kind of text a model is weak at.
|
| 70 |
+
- The whole pipeline is included (`code/`) and every filter decision is documented below.
|
| 71 |
+
|
| 72 |
+
## Subsets
|
| 73 |
+
|
| 74 |
+
| Subset | Train words | Train docs | ~Tokens (BPE) |
|
| 75 |
+
|---|---|---|---|
|
| 76 |
+
| `1M` | 1.0M | 4,199 | ~1.3M |
|
| 77 |
+
| `10M` | 9.95M | 42,742 | ~12.9M |
|
| 78 |
+
| `100M` | 99.51M | 336,124 | ~129.4M |
|
| 79 |
+
|
| 80 |
+
Words are whitespace-separated words in `train`. Each subset also has about 0.5% of the documents per genre held out as `validation_<genre>` splits
|
| 81 |
+
(at least one document per genre; the `1M` validation splits are therefore tiny, a few dozen documents, and noisy).
|
| 82 |
+
|
| 83 |
+
```python
|
| 84 |
+
from datasets import load_dataset
|
| 85 |
+
ds = load_dataset("jncarlo/tinymix", "10M") # splits: train, validation_narrative, validation_dialogue, ...
|
| 86 |
+
print(ds["train"][0]) # {'text': ..., 'source': 'tinystories', 'genre': 'narrative'}
|
| 87 |
+
```
|
| 88 |
+
|
| 89 |
+
Columns: `text` (str), `source` (key of the table below), `genre` (`narrative`, `dialogue`, `expository`, `books`, `poetry`, `news`).
|
| 90 |
+
Documents are independent; join them with an end-of-text token when training. Dialogues are formatted one turn per line
|
| 91 |
+
(`Mia: ...`, or `User:` / `Assistant:` for assistant chats).
|
| 92 |
+
|
| 93 |
+
## Composition
|
| 94 |
+
|
| 95 |
+
Share of training words by genre:
|
| 96 |
+
|
| 97 |
+
| Genre | 1M | 10M | 100M |
|
| 98 |
+
|---|---|---|---|
|
| 99 |
+
| narrative | 37.9% | 35.0% | 21.0% |
|
| 100 |
+
| dialogue | 22.0% | 25.0% | 21.0% |
|
| 101 |
+
| expository | 40.1% | 40.0% | 48.0% |
|
| 102 |
+
| books | 0% | 0% | 8.0% |
|
| 103 |
+
| poetry | 0% | 0% | 1.0% |
|
| 104 |
+
| news | 0% | 0% | 1.0% |
|
| 105 |
+
|
| 106 |
+
Million words per source:
|
| 107 |
+
|
| 108 |
+
| Source | 1M (M words) | 10M (M words) | 100M (M words) |
|
| 109 |
+
|---|---|---|---|
|
| 110 |
+
| TinyStories V2 (GPT-4 split) | 0.189 | 1.492 | 7.963 |
|
| 111 |
+
| SimpleStories | 0.189 | 1.99 | 7.959 |
|
| 112 |
+
| Cosmopedia v0.1, `stories` | 0 | 0 | 4.974 |
|
| 113 |
+
| TinyDialogues | 0.12 | 1.493 | 6.969 |
|
| 114 |
+
| SODA | 0.1 | 0.994 | 8.952 |
|
| 115 |
+
| UltraChat 200k | 0 | 0 | 3.984 |
|
| 116 |
+
| SmolTalk (everyday-conversations, smol-summarize, smol-rewrite) | 0 | 0 | 0.995 |
|
| 117 |
+
| TinyHelen Leaner-Pretrain (book, conversation, textbook, web, wiki) | 0.15 | 1.491 | 9.948 |
|
| 118 |
+
| Simple English Wikipedia (20231101.simple) | 0.1 | 0.516 | 9.949 |
|
| 119 |
+
| Cosmopedia v2 (SmolLM-Corpus) | 0.1 | 1.113 | 11.952 |
|
| 120 |
+
| Cosmopedia v0.1 (wikihow, khanacademy, openstax) | 0.051 | 0.383 | 2.985 |
|
| 121 |
+
| tiny-textbooks (`textbook` column) | 0 | 0.479 | 3.98 |
|
| 122 |
+
| FineWeb-Edu (sample-10BT, score >= 4) | 0 | 0 | 8.952 |
|
| 123 |
+
| Project Gutenberg (English books) | 0 | 0 | 7.961 |
|
| 124 |
+
| Gutenberg Poetry Corpus | 0 | 0 | 0.995 |
|
| 125 |
+
| CNN/DailyMail 3.0.0 | 0 | 0 | 0.995 |
|
| 126 |
+
|
| 127 |
+
The 10M and 100M subsets reached their targets (10.0M and 100.0M words including the validation splits). In the 10M build,
|
| 128 |
+
sources that could not fill their quota with the filters (Simple English Wikipedia, Cosmopedia v0.1) were topped up by
|
| 129 |
+
same-genre substitute sources (tiny-textbooks, Cosmopedia v2), capped per source; that is why `tiny-textbooks` appears
|
| 130 |
+
there. The 1M subset keeps the 10M proportions except that Cosmopedia stories (not in the 10M mix) were replaced by TinyStories and SimpleStories.
|
| 131 |
+
|
| 132 |
+
## How it was built
|
| 133 |
+
|
| 134 |
+
1. **Normalisation**: `ftfy`, NFKC, ASCII quotes and dashes, whitespace; URLs, e-mails, phone numbers, code, LaTeX and HTML removed.
|
| 135 |
+
2. **Language**: fastText `lid.176`, English with confidence >= 0.9 (skipped for TinyDialogues, whose one-word toddler turns confuse it).
|
| 136 |
+
3. **Quality heuristics**: Gopher-style rules (mean word length, symbol and digit ratio, repeated lines, repeated bigrams, stop words), minimum length, complete ending (drops truncated generations), a short list of LLM tics (`delve`, `tapestry`, ...) and an explicit-content word list.
|
| 137 |
+
4. **Anti-template**: no 4-word opening may exceed 0.5% of a source's documents.
|
| 138 |
+
5. **Deduplication across all sources**: exact hash and MinHash LSH (5-word shingles, Jaccard >= 0.8).
|
| 139 |
+
6. **Complexity control by size** (vocabulary coverage = share of lowercase tokens inside the top-N English words of `wordfreq`):
|
| 140 |
+
|
| 141 |
+
| | 10M | 100M |
|
| 142 |
+
|---|---|---|
|
| 143 |
+
| Reference vocabulary | top 8k | top 15k |
|
| 144 |
+
| Min vocabulary coverage | 90% | 90% |
|
| 145 |
+
| Max Flesch-Kincaid grade (narrative, news, books) | 5 | 8 |
|
| 146 |
+
| Max Flesch-Kincaid grade (expository) | 9 | 10 |
|
| 147 |
+
| Document length | 80 to 600 words | 80 to 1200 words |
|
| 148 |
+
| Dialogue: mean words per turn | <= 25 | <= 40 |
|
| 149 |
+
|
| 150 |
+
Assistant chats (`User:` / `Assistant:`) may average up to 4 times the per-turn cap and need at least 2 turns (named dialogues need 4),
|
| 151 |
+
because summarisation and rewrite tasks have 2 turns and assistant answers are long.
|
| 152 |
+
|
| 153 |
+
Documents longer than the maximum are split at paragraph borders. The 1M subset is a prefix slice of the 10M build; 10M and 100M were built separately with the same recipe, so they are **not** guaranteed to be subsets of each other (a later fix to the end-of-document check, which wrongly rejected Wikipedia articles that end in a heading, was only applied to sources rebuilt for the 100M run).
|
| 154 |
+
7. **Run safety nets** (used to build the larger subsets without manual tuning): a probe per source picks a starting filter strictness, strictness relaxes in up to three logged steps when yield is too low, and shortfalls go to substitute sources.
|
| 155 |
+
|
| 156 |
+
Thresholds were calibrated on real data: the first values (95% coverage in the top 5k words, Flesch-Kincaid <= 5 for everything)
|
| 157 |
+
accepted only 3% of SimpleStories and 0.6% of Cosmopedia v2, so they were relaxed as described above.
|
| 158 |
+
|
| 159 |
+
## Sources and licensing
|
| 160 |
+
|
| 161 |
+
| Source | License | Role |
|
| 162 |
+
|---|---|---|
|
| 163 |
+
| [TinyStories V2 (GPT-4 split)](https://huggingface.co/datasets/roneneldan/TinyStories) | CDLA-Sharing-1.0 | child-level stories |
|
| 164 |
+
| [SimpleStories](https://huggingface.co/datasets/SimpleStories/SimpleStories) | MIT | varied simple stories |
|
| 165 |
+
| [Cosmopedia v0.1, `stories`](https://huggingface.co/datasets/HuggingFaceTB/cosmopedia) | Apache-2.0 | everyday stories |
|
| 166 |
+
| [TinyDialogues](https://huggingface.co/datasets/styfeng/TinyDialogues) | MIT | child-directed dialogue |
|
| 167 |
+
| [SODA](https://huggingface.co/datasets/allenai/soda) | CC-BY-4.0 | social dialogue |
|
| 168 |
+
| [UltraChat 200k](https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k) | MIT | user/assistant chat |
|
| 169 |
+
| [SmolTalk (everyday-conversations, smol-summarize, smol-rewrite)](https://huggingface.co/datasets/HuggingFaceTB/smoltalk) | Apache-2.0 | chat, summaries, rewrites |
|
| 170 |
+
| [TinyHelen Leaner-Pretrain (book, conversation, textbook, web, wiki)](https://huggingface.co/datasets/EmpathYang/TinyHelen) | MIT | simplified real text |
|
| 171 |
+
| [Simple English Wikipedia (20231101.simple)](https://huggingface.co/datasets/wikimedia/wikipedia) | CC-BY-SA-3.0 / GFDL | encyclopedic, simple |
|
| 172 |
+
| [Cosmopedia v2 (SmolLM-Corpus)](https://huggingface.co/datasets/HuggingFaceTB/smollm-corpus) | Apache-2.0 (corpus: ODC-By) | synthetic textbooks, blogs, stories |
|
| 173 |
+
| [Cosmopedia v0.1 (wikihow, khanacademy, openstax)](https://huggingface.co/datasets/HuggingFaceTB/cosmopedia) | Apache-2.0 | procedural and didactic |
|
| 174 |
+
| [tiny-textbooks (`textbook` column)](https://huggingface.co/datasets/nampdn-ai/tiny-textbooks) | Apache-2.0 | short synthetic lessons |
|
| 175 |
+
| [FineWeb-Edu (sample-10BT, score >= 4)](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu) | ODC-By-1.0 | educational web |
|
| 176 |
+
| [Project Gutenberg (English books)](https://huggingface.co/datasets/manu/project_gutenberg) | Project Gutenberg License (public domain in the US) | classic books |
|
| 177 |
+
| [Gutenberg Poetry Corpus](https://huggingface.co/datasets/biglam/gutenberg-poetry-corpus) | CC0-1.0 | poetry stanzas |
|
| 178 |
+
| [CNN/DailyMail 3.0.0](https://huggingface.co/datasets/abisee/cnn_dailymail) | Apache-2.0 (articles are (c) CNN / Daily Mail) | news |
|
| 179 |
+
|
| 180 |
+
**Licensing and attribution.** This repository is a derivative compilation of the datasets above; each document keeps the
|
| 181 |
+
license of its `source`. Obligations that matter when redistributing or reusing it:
|
| 182 |
+
|
| 183 |
+
- Attribution is required for CC-BY / ODC-By / CDLA sources (listed above).
|
| 184 |
+
- The Wikipedia part (`simplewiki`) is **CC-BY-SA-3.0 / GFDL (share-alike)**. If you need a share-alike-free corpus, filter out `source == "simplewiki"`.
|
| 185 |
+
- Cosmopedia, SmolTalk, UltraChat, TinyStories, SimpleStories, TinyHelen and tiny-textbooks contain **model-generated text**; check the terms of the generating model (GPT-4, GPT-4o-mini, Mixtral and others) for your use case.
|
| 186 |
+
- CNN/DailyMail articles and FineWeb-Edu web pages are third-party content; they are included under the dataset licenses above, which do not transfer rights from the original publishers.
|
| 187 |
+
- Project Gutenberg texts had the Project Gutenberg header and footer removed; they are public domain in the United States, check your jurisdiction.
|
| 188 |
+
|
| 189 |
+
No personal data was added; e-mails, phone numbers and URLs were removed, but residual personal information in web-derived sources cannot be ruled out.
|
| 190 |
+
|
| 191 |
+
## Sanity-check training runs
|
| 192 |
+
|
| 193 |
+
Small GPTs (nanoGPT-style, own byte-level BPE trained on each training set) trained for the same number of steps within each size (17M parameters for 10M words, 46M for 100M words), scored in **bits per byte** (lower is better, comparable across tokenizers) on the held-out validation splits of **this** dataset. Baselines take the first 10M / 100M words of a single source with only `ftfy` cleaning. They are out of distribution on the genres they do not cover, so the table shows how much genre coverage matters, not a leaderboard. One seed, no tuning; the 10M and 100M rows are not comparable with each other.
|
| 194 |
+
|
| 195 |
+
| Training data | Params | Train tokens | narrative | dialogue | expository | books | news | poetry |
|
| 196 |
+
|---|---|---|---|---|---|---|---|---|
|
| 197 |
+
| tinymix 10M | 17.41M | 14.11M | 0.833 | 0.792 | 1.278 | n/a | n/a | n/a |
|
| 198 |
+
| TinyStories only, 10M words | 17.41M | 12.05M | 1.111 | 2.512 | 2.708 | n/a | n/a | n/a |
|
| 199 |
+
| FineWeb-Edu only, 10M words | 17.41M | 15.71M | 1.639 | 1.769 | 1.565 | n/a | n/a | n/a |
|
| 200 |
+
| tinymix 100M | 46.29M | 137.02M | 0.731 | 0.721 | 1.043 | 1.401 | 1.456 | 1.843 |
|
| 201 |
+
|
| 202 |
+
## Limitations
|
| 203 |
+
|
| 204 |
+
- **Not done**: perplexity filtering with an n-gram LM, LLM-judge coherence scoring, semantic deduplication, cluster caps and benchmark decontamination. Expect some residual near-duplicates and some low-quality documents.
|
| 205 |
+
- Some `tinyhelen` chunks start mid-thought, because the simplified web and book documents are chunks of longer texts.
|
| 206 |
+
- Synthetic sources share model-specific style and topics (names, morals, classroom settings). The anti-template cap reduces this but does not remove it.
|
| 207 |
+
- The toxicity filter is a short word list, not a classifier. Chat and web sources can still contain mildly sensitive content.
|
| 208 |
+
- English only, mostly simple modern language; it is a poor fit for models meant to handle technical, legal or archaic text.
|
| 209 |
+
- Word counts are whitespace words; token counts depend on your tokenizer.
|
| 210 |
+
|
| 211 |
+
## Reproduce
|
| 212 |
+
|
| 213 |
+
`code/` contains the build scripts (`mix.py` quotas and tiers, `sources.py` adapters, `build.py` pipeline, `assemble.py`,
|
| 214 |
+
`publish.py`, `train.py` for the sanity-check runs) and `requirements.txt`. Building a subset: `python build.py 10M`,
|
| 215 |
+
then `python assemble.py 10M`. The build streams the source datasets from the Hub.
|
| 216 |
+
|
| 217 |
+
## Acknowledgements
|
| 218 |
+
|
| 219 |
+
Built on the work behind TinyStories, SimpleStories, TinyDialogues, SODA, UltraChat, SmolTalk and SmolLM-Corpus
|
| 220 |
+
(Cosmopedia, FineWeb-Edu), TinyHelen, tiny-textbooks, Project Gutenberg, the Gutenberg Poetry Corpus, Wikipedia and
|
| 221 |
+
CNN/DailyMail. Please cite the original datasets when you use this one.
|
code/assemble.py
ADDED
|
@@ -0,0 +1,87 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Assemble a size: python assemble.py 10M
|
| 2 |
+
|
| 3 |
+
Shuffles documents with a fixed seed, holds out 0.5% per genre as validation,
|
| 4 |
+
writes out/<size>/train.txt (docs separated by <|endoftext|>) and valid_<genre>.txt.
|
| 5 |
+
"""
|
| 6 |
+
import json
|
| 7 |
+
import random
|
| 8 |
+
import sys
|
| 9 |
+
from collections import defaultdict
|
| 10 |
+
from pathlib import Path
|
| 11 |
+
|
| 12 |
+
ROOT = Path(__file__).resolve().parent.parent
|
| 13 |
+
EOT = "<|endoftext|>"
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
# 1M is a prefix slice of the 10M build (million words per source)
|
| 17 |
+
QUOTAS_1M = {"tinystories": 0.19, "simplestories": 0.19, "tinydialogues": 0.12, "soda": 0.10,
|
| 18 |
+
"tinyhelen": 0.15, "simplewiki": 0.10, "cosmo_v2": 0.10, "cosmo_v01": 0.05}
|
| 19 |
+
|
| 20 |
+
|
| 21 |
+
def load(size):
|
| 22 |
+
docs = []
|
| 23 |
+
if size == "1M":
|
| 24 |
+
for name, q in QUOTAS_1M.items():
|
| 25 |
+
n = 0
|
| 26 |
+
for l in (ROOT / "work" / "10M" / f"{name}.jsonl").open(encoding="utf-8"):
|
| 27 |
+
d = json.loads(l)
|
| 28 |
+
if n >= q * 1e6:
|
| 29 |
+
break
|
| 30 |
+
n += len(d["text"].split())
|
| 31 |
+
docs.append(d)
|
| 32 |
+
return docs
|
| 33 |
+
for p in sorted((ROOT / "work" / size).glob("*.jsonl")):
|
| 34 |
+
docs += [json.loads(l) for l in p.open(encoding="utf-8")]
|
| 35 |
+
return docs
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
def main():
|
| 39 |
+
size = sys.argv[1]
|
| 40 |
+
docs = load(size)
|
| 41 |
+
random.Random(0).shuffle(docs)
|
| 42 |
+
out = ROOT / "out" / size
|
| 43 |
+
out.mkdir(parents=True, exist_ok=True)
|
| 44 |
+
valid, held = defaultdict(list), defaultdict(int)
|
| 45 |
+
n_genre = defaultdict(int)
|
| 46 |
+
for d in docs:
|
| 47 |
+
n_genre[d["genre"]] += 1
|
| 48 |
+
train, words, by_genre, by_source = [], 0, defaultdict(int), defaultdict(int)
|
| 49 |
+
valid_docs = defaultdict(list)
|
| 50 |
+
for d in docs:
|
| 51 |
+
g = d["genre"]
|
| 52 |
+
if held[g] < max(1, n_genre[g] // 200):
|
| 53 |
+
held[g] += 1
|
| 54 |
+
valid[g].append(d["text"])
|
| 55 |
+
valid_docs[g].append(d)
|
| 56 |
+
else:
|
| 57 |
+
train.append(d)
|
| 58 |
+
n = len(d["text"].split())
|
| 59 |
+
by_genre[g] += n
|
| 60 |
+
by_source[d["src"].split(".")[0]] += n
|
| 61 |
+
words += n
|
| 62 |
+
sep = f"\n{EOT}\n"
|
| 63 |
+
(out / "train.txt").write_text(sep.join(d["text"] for d in train) + sep, encoding="utf-8")
|
| 64 |
+
for g, ts in valid.items():
|
| 65 |
+
(out / f"valid_{g}.txt").write_text(sep.join(ts) + sep, encoding="utf-8")
|
| 66 |
+
# parquet with provenance, for the Hugging Face release
|
| 67 |
+
import pyarrow as pa
|
| 68 |
+
import pyarrow.parquet as pq
|
| 69 |
+
pq_dir = out / "parquet"
|
| 70 |
+
pq_dir.mkdir(exist_ok=True)
|
| 71 |
+
|
| 72 |
+
def write(name, ds):
|
| 73 |
+
pq.write_table(pa.table({"text": [d["text"] for d in ds], "source": [d["src"].split(".")[0] for d in ds],
|
| 74 |
+
"genre": [d["genre"] for d in ds]}), pq_dir / f"{name}.parquet", compression="zstd")
|
| 75 |
+
|
| 76 |
+
write("train", train)
|
| 77 |
+
for g, ds in valid_docs.items():
|
| 78 |
+
write(f"validation_{g}", ds)
|
| 79 |
+
summary = {"size": size, "train_docs": len(train), "train_words_M": round(words / 1e6, 2),
|
| 80 |
+
"genre_share_%": {g: round(100 * w / words, 1) for g, w in by_genre.items()},
|
| 81 |
+
"source_words_M": {s: round(w / 1e6, 3) for s, w in sorted(by_source.items())}}
|
| 82 |
+
(out / "summary.json").write_text(json.dumps(summary, indent=1))
|
| 83 |
+
print(json.dumps(summary, indent=1))
|
| 84 |
+
|
| 85 |
+
|
| 86 |
+
if __name__ == "__main__":
|
| 87 |
+
main()
|
code/baselines.py
ADDED
|
@@ -0,0 +1,71 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Single-source baselines of the same size, for comparison with the mix.
|
| 2 |
+
|
| 3 |
+
python baselines.py 10M -> out/base_tinystories_10M, out/base_fineweb_edu_10M
|
| 4 |
+
Light cleaning only (ftfy + whitespace), no mix, no complexity or dedup filters:
|
| 5 |
+
this is what someone would get by just taking the dataset.
|
| 6 |
+
"""
|
| 7 |
+
import os
|
| 8 |
+
import sys
|
| 9 |
+
from pathlib import Path
|
| 10 |
+
|
| 11 |
+
ROOT = Path(__file__).resolve().parent.parent
|
| 12 |
+
_env = ROOT / ".env"
|
| 13 |
+
if _env.exists():
|
| 14 |
+
for _l in _env.read_text(encoding="utf-8").splitlines():
|
| 15 |
+
if "=" in _l and not _l.lstrip().startswith("#"):
|
| 16 |
+
_k, _v = _l.split("=", 1)
|
| 17 |
+
os.environ.setdefault(_k.strip(), _v.strip().strip("'\""))
|
| 18 |
+
|
| 19 |
+
import ftfy
|
| 20 |
+
from datasets import load_dataset
|
| 21 |
+
from huggingface_hub import HfFileSystem
|
| 22 |
+
|
| 23 |
+
EOT = "<|endoftext|>"
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
def tinystories():
|
| 27 |
+
buf = []
|
| 28 |
+
with HfFileSystem().open("datasets/roneneldan/TinyStories/TinyStoriesV2-GPT4-train.txt", "r",
|
| 29 |
+
encoding="utf-8", errors="replace") as f:
|
| 30 |
+
for line in f:
|
| 31 |
+
if EOT in line:
|
| 32 |
+
if buf:
|
| 33 |
+
yield " ".join(buf)
|
| 34 |
+
buf = []
|
| 35 |
+
else:
|
| 36 |
+
buf.append(line.strip())
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
def fineweb_edu():
|
| 40 |
+
for r in load_dataset("HuggingFaceFW/fineweb-edu", "sample-10BT", split="train", streaming=True):
|
| 41 |
+
if r["int_score"] >= 4:
|
| 42 |
+
yield r["text"]
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
SOURCES = {"tinystories": tinystories, "fineweb_edu": fineweb_edu}
|
| 46 |
+
|
| 47 |
+
|
| 48 |
+
def main():
|
| 49 |
+
size = sys.argv[1]
|
| 50 |
+
target = {"10M": 10e6, "100M": 100e6}[size]
|
| 51 |
+
for name, gen in SOURCES.items():
|
| 52 |
+
if len(sys.argv) > 2 and name not in sys.argv[2:]:
|
| 53 |
+
continue
|
| 54 |
+
out = ROOT / "out" / f"base_{name}_{size}"
|
| 55 |
+
out.mkdir(parents=True, exist_ok=True)
|
| 56 |
+
words, n = 0, 0
|
| 57 |
+
with (out / "train.txt").open("w", encoding="utf-8") as f:
|
| 58 |
+
for text in gen():
|
| 59 |
+
text = ftfy.fix_text(text).strip()
|
| 60 |
+
if not text:
|
| 61 |
+
continue
|
| 62 |
+
f.write(text + f"\n{EOT}\n")
|
| 63 |
+
words += len(text.split())
|
| 64 |
+
n += 1
|
| 65 |
+
if words >= target:
|
| 66 |
+
break
|
| 67 |
+
print(f"{name} {size}: {words/1e6:.2f}M words, {n} docs", flush=True)
|
| 68 |
+
|
| 69 |
+
|
| 70 |
+
if __name__ == "__main__":
|
| 71 |
+
main()
|
code/build.py
ADDED
|
@@ -0,0 +1,378 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Build one size of the dataset: python build.py 10M [source ...]
|
| 2 |
+
|
| 3 |
+
Per source: stream -> clean -> filter -> dedup until the word quota is met.
|
| 4 |
+
Output: work/<size>/<source>.jsonl, then assemble.py slices train/valid.
|
| 5 |
+
"""
|
| 6 |
+
import hashlib
|
| 7 |
+
import itertools
|
| 8 |
+
import json
|
| 9 |
+
import os
|
| 10 |
+
import re
|
| 11 |
+
import sys
|
| 12 |
+
import time
|
| 13 |
+
import unicodedata
|
| 14 |
+
from collections import Counter
|
| 15 |
+
from pathlib import Path
|
| 16 |
+
|
| 17 |
+
ROOT = Path(__file__).resolve().parent.parent
|
| 18 |
+
# HF_TOKEN lives in ../.env (never printed); must be set before importing datasets
|
| 19 |
+
_env = ROOT / ".env"
|
| 20 |
+
if _env.exists():
|
| 21 |
+
for _l in _env.read_text(encoding="utf-8").splitlines():
|
| 22 |
+
if "=" in _l and not _l.lstrip().startswith("#"):
|
| 23 |
+
_k, _v = _l.split("=", 1)
|
| 24 |
+
os.environ.setdefault(_k.strip(), _v.strip().strip("'\""))
|
| 25 |
+
|
| 26 |
+
import fasttext
|
| 27 |
+
import ftfy
|
| 28 |
+
from datasketch import MinHash, MinHashLSH
|
| 29 |
+
from wordfreq import top_n_list
|
| 30 |
+
|
| 31 |
+
import mix
|
| 32 |
+
import sources
|
| 33 |
+
|
| 34 |
+
QUOTES = str.maketrans({"‘": "'", "’": "'", "“": '"', "”": '"',
|
| 35 |
+
"–": "-", "—": " - ", "…": "...", " ": " "})
|
| 36 |
+
RE_URL = re.compile(r"https?://\S+|www\.\S+")
|
| 37 |
+
RE_EMAIL = re.compile(r"\S+@\S+\.\S+")
|
| 38 |
+
RE_PHONE = re.compile(r"\+?\d[\d\s().-]{8,}\d")
|
| 39 |
+
RE_HTML = re.compile(r"<[^>]{1,100}>")
|
| 40 |
+
RE_CODE = re.compile(r"```.*?```", re.S)
|
| 41 |
+
RE_HEAD = re.compile(r"^#{1,6}\s*", re.M)
|
| 42 |
+
RE_LATEX = re.compile(r"\$\$.*?\$\$|\\\(.*?\\\)|\\\[.*?\\\]", re.S)
|
| 43 |
+
LLM_TICS = re.compile(r"\bas an ai\b|\bdelve\b|\btapestry\b|\bcertainly!|\bin conclusion,|"
|
| 44 |
+
r"\bit'?s important to note\b|\bbustling\b|\bas a language model\b", re.I)
|
| 45 |
+
# deliberately small: only terms that are unambiguous in any context
|
| 46 |
+
BAD = re.compile(r"\b(fuck\w*|shit\w*|bitch\w*|cunt|porn\w*|nigg\w+|rape[sd]?|raping|"
|
| 47 |
+
r"cock|dick|pussy|slut|whore|orgasm\w*|erotic\w*|blowjob|masturbat\w*)\b", re.I)
|
| 48 |
+
STOP = {"the", "and", "to", "of", "a", "in", "that", "is", "it", "for"}
|
| 49 |
+
END_OK = ('.', '!', '?', '"', "'", ")")
|
| 50 |
+
|
| 51 |
+
lid = fasttext.load_model(str(ROOT / "work" / "lid.176.ftz"))
|
| 52 |
+
|
| 53 |
+
|
| 54 |
+
def normalize(t):
|
| 55 |
+
t = ftfy.fix_text(t).translate(QUOTES)
|
| 56 |
+
t = unicodedata.normalize("NFKC", t)
|
| 57 |
+
for rx in (RE_CODE, RE_LATEX, RE_URL, RE_EMAIL, RE_HTML):
|
| 58 |
+
t = rx.sub(" ", t)
|
| 59 |
+
t = RE_PHONE.sub(" ", t)
|
| 60 |
+
t = RE_HEAD.sub("", t)
|
| 61 |
+
t = re.sub(r"[ \t]+", " ", t)
|
| 62 |
+
t = re.sub(r" ?\n ?", "\n", t)
|
| 63 |
+
t = re.sub(r"\n{3,}", "\n\n", t)
|
| 64 |
+
return t.strip()
|
| 65 |
+
|
| 66 |
+
|
| 67 |
+
def syllables(w):
|
| 68 |
+
n = len(re.findall(r"[aeiouy]+", w))
|
| 69 |
+
if w.endswith("e") and n > 1 and not w.endswith("le"):
|
| 70 |
+
n -= 1
|
| 71 |
+
return max(n, 1)
|
| 72 |
+
|
| 73 |
+
|
| 74 |
+
def fk_grade(text):
|
| 75 |
+
words = re.findall(r"[A-Za-z']+", text)
|
| 76 |
+
sents = max(len(re.findall(r"[.!?]+(?:\s|$)", text)), 1)
|
| 77 |
+
if not words:
|
| 78 |
+
return 99
|
| 79 |
+
syl = sum(syllables(w.lower()) for w in words)
|
| 80 |
+
return 0.39 * len(words) / sents + 11.8 * syl / len(words) - 15.59
|
| 81 |
+
|
| 82 |
+
|
| 83 |
+
def vocab_coverage(text, vocab):
|
| 84 |
+
# capitalised words (names, sentence starts) never match, so they are ignored
|
| 85 |
+
toks = re.findall(r"\b[a-z']+\b", text)
|
| 86 |
+
return sum(t in vocab for t in toks) / len(toks) if toks else 0
|
| 87 |
+
|
| 88 |
+
|
| 89 |
+
def gopher_ok(text):
|
| 90 |
+
words = text.split()
|
| 91 |
+
if not words:
|
| 92 |
+
return False
|
| 93 |
+
if not 3 <= sum(map(len, words)) / len(words) <= 10:
|
| 94 |
+
return False
|
| 95 |
+
if sum(not c.isalnum() and not c.isspace() and c not in ".,;:!?'\"()-" for c in text) / len(text) > 0.1:
|
| 96 |
+
return False
|
| 97 |
+
if sum(c.isdigit() for c in text) / len(text) > 0.1:
|
| 98 |
+
return False
|
| 99 |
+
lines = [l for l in text.split("\n") if l.strip()]
|
| 100 |
+
if len(lines) > 3 and 1 - len(set(lines)) / len(lines) > 0.3:
|
| 101 |
+
return False
|
| 102 |
+
if sum(w.lower().strip(".,!?\"'") in STOP for w in words) < 2:
|
| 103 |
+
return False
|
| 104 |
+
bi = Counter(zip(words, words[1:]))
|
| 105 |
+
return not bi or bi.most_common(1)[0][1] / max(len(words) - 1, 1) <= 0.2
|
| 106 |
+
|
| 107 |
+
|
| 108 |
+
def is_english(text):
|
| 109 |
+
# lid.f.predict avoids fasttext's numpy-1 only wrapper; returns [(prob, label)]
|
| 110 |
+
pred = lid.f.predict(text[:1000].replace("\n", " ") + "\n", 1, 0.0, "strict")
|
| 111 |
+
return bool(pred) and pred[0][1] == "__label__en" and pred[0][0] >= 0.9
|
| 112 |
+
|
| 113 |
+
|
| 114 |
+
def split_long(text, title, doc_max):
|
| 115 |
+
"""Split at paragraph borders; the title is repeated so each chunk stays self-contained."""
|
| 116 |
+
paras, chunks, cur, n = text.split("\n"), [], [], 0
|
| 117 |
+
for p in paras:
|
| 118 |
+
w = len(p.split())
|
| 119 |
+
if cur and n + w > doc_max:
|
| 120 |
+
chunks.append("\n".join(cur))
|
| 121 |
+
cur, n = [], 0
|
| 122 |
+
cur.append(p)
|
| 123 |
+
n += w
|
| 124 |
+
if cur:
|
| 125 |
+
chunks.append("\n".join(cur))
|
| 126 |
+
return [(f"{title}\n{c}" if title else c) for c in chunks if len(c.split()) >= mix.DOC_MIN]
|
| 127 |
+
|
| 128 |
+
|
| 129 |
+
class Dedup:
|
| 130 |
+
def __init__(self):
|
| 131 |
+
self.exact, self.lsh, self.n = set(), MinHashLSH(threshold=0.8, num_perm=64), 0
|
| 132 |
+
|
| 133 |
+
def seen(self, text):
|
| 134 |
+
h = hashlib.md5(re.sub(r"\W+", "", text.lower()).encode()).hexdigest()
|
| 135 |
+
if h in self.exact:
|
| 136 |
+
return True
|
| 137 |
+
self.exact.add(h)
|
| 138 |
+
w = text.lower().split()
|
| 139 |
+
mh = MinHash(num_perm=64)
|
| 140 |
+
mh.update_batch([" ".join(w[i:i + 5]).encode() for i in range(max(len(w) - 4, 1))])
|
| 141 |
+
if self.lsh.query(mh):
|
| 142 |
+
return True
|
| 143 |
+
self.lsh.insert(str(self.n), mh)
|
| 144 |
+
self.n += 1
|
| 145 |
+
return False
|
| 146 |
+
|
| 147 |
+
|
| 148 |
+
def trim_tail(text):
|
| 149 |
+
"""Drop trailing headings and list stubs ("Other websites", "References") so a complete
|
| 150 |
+
article is not rejected as truncated; real truncation leaves a long unfinished last line."""
|
| 151 |
+
lines = text.split("\n")
|
| 152 |
+
while len(lines) > 1 and (not lines[-1].strip() or
|
| 153 |
+
(len(lines[-1].split()) < 12 and not lines[-1].rstrip().endswith(END_OK))):
|
| 154 |
+
lines.pop()
|
| 155 |
+
return "\n".join(lines).strip()
|
| 156 |
+
|
| 157 |
+
|
| 158 |
+
def process(doc, genre, tier, vocab, src_name, cov):
|
| 159 |
+
"""Return (clean documents, None) or ([], rejection reason)."""
|
| 160 |
+
if "turns" in doc:
|
| 161 |
+
turns = [(s.strip(), normalize(t)) for s, t in doc["turns"]]
|
| 162 |
+
turns = [(s, t.replace("\n", " ")) for s, t in turns if t]
|
| 163 |
+
# assistant chats (User/Assistant) include 2-turn tasks (summarise, rewrite) and long answers,
|
| 164 |
+
# so they get a lower turn minimum and a looser per-turn cap than named dialogues
|
| 165 |
+
chat = all(s in ("User", "Assistant") for s, _ in turns)
|
| 166 |
+
if len(turns) < (2 if chat else 4) or any(len(s) > 20 for s, _ in turns):
|
| 167 |
+
return [], "dialogue_shape"
|
| 168 |
+
if sum(len(t.split()) for _, t in turns) / len(turns) > tier["turn"] * (mix.CHAT_TURN_MULT if chat else 1):
|
| 169 |
+
return [], "turn_length"
|
| 170 |
+
text = "\n".join(f"{s}: {t}" for s, t in turns)
|
| 171 |
+
title = None
|
| 172 |
+
else:
|
| 173 |
+
text, title = normalize(doc["text"]), doc.get("title")
|
| 174 |
+
if "�" in text:
|
| 175 |
+
return [], "mojibake"
|
| 176 |
+
if genre != "poetry":
|
| 177 |
+
text = trim_tail(text)
|
| 178 |
+
nw = len(text.split())
|
| 179 |
+
if genre == "poetry":
|
| 180 |
+
if nw < 20:
|
| 181 |
+
return [], "short"
|
| 182 |
+
return ([text], None) if is_english(text) else ([], "lang")
|
| 183 |
+
if nw < mix.DOC_MIN:
|
| 184 |
+
return [], "short"
|
| 185 |
+
if len(LLM_TICS.findall(text)) > 2:
|
| 186 |
+
return [], "llm_tics"
|
| 187 |
+
if BAD.search(text):
|
| 188 |
+
return [], "explicit"
|
| 189 |
+
if src_name not in mix.LID_SKIP and not is_english(text):
|
| 190 |
+
return [], "lang"
|
| 191 |
+
if not text.rstrip().endswith(END_OK):
|
| 192 |
+
return [], "truncated"
|
| 193 |
+
if not gopher_ok(text):
|
| 194 |
+
return [], "gopher"
|
| 195 |
+
if genre != "dialogue":
|
| 196 |
+
cap = tier["fk_exp"] if genre == "expository" else tier["fk"]
|
| 197 |
+
if fk_grade(text) > cap and src_name not in ("simplewiki",):
|
| 198 |
+
return [], "fk"
|
| 199 |
+
if vocab_coverage(text, vocab) < cov:
|
| 200 |
+
return [], "vocab"
|
| 201 |
+
docs = split_long(text, title, tier["doc_max"]) if nw > tier["doc_max"] else [f"{title}\n{text}" if title else text]
|
| 202 |
+
return docs, (None if docs else "short")
|
| 203 |
+
|
| 204 |
+
|
| 205 |
+
def doc_words(doc):
|
| 206 |
+
return len(doc["text"].split()) if "text" in doc else sum(len(t.split()) for _, t in doc["turns"])
|
| 207 |
+
|
| 208 |
+
|
| 209 |
+
def at_step(tier, step):
|
| 210 |
+
"""Tier and vocabulary-coverage threshold at a relaxation-ladder step."""
|
| 211 |
+
dfk, dcov = mix.LADDER[step]
|
| 212 |
+
return dict(tier, fk=tier["fk"] + dfk, fk_exp=tier["fk_exp"] + dfk), mix.VOCAB_COVERAGE + dcov
|
| 213 |
+
|
| 214 |
+
|
| 215 |
+
def probe(name, size, genre, tier, vocab, quota):
|
| 216 |
+
"""Mechanism 1: sample the first docs and pick the first ladder step whose projected
|
| 217 |
+
scan fits comfortably in the budget (yield >= 2/MAX_SCAN). Returns (step, report)."""
|
| 218 |
+
docs = list(itertools.islice(sources.REGISTRY[name](size), mix.PROBE_DOCS))
|
| 219 |
+
scanned = sum(doc_words(d) for d in docs) or 1
|
| 220 |
+
report = []
|
| 221 |
+
for step in range(len(mix.LADDER)):
|
| 222 |
+
t, cov = at_step(tier, step)
|
| 223 |
+
rej, kept = Counter(), 0
|
| 224 |
+
for d in docs:
|
| 225 |
+
out, why = process(d, genre, t, vocab, name, cov)
|
| 226 |
+
kept += sum(len(x.split()) for x in out)
|
| 227 |
+
if why:
|
| 228 |
+
rej[why] += 1
|
| 229 |
+
yield_ = kept / scanned
|
| 230 |
+
report.append(dict(step=step, yield_pct=round(100 * yield_, 1), top_rejections=rej.most_common(3)))
|
| 231 |
+
if yield_ >= 2 / mix.MAX_SCAN:
|
| 232 |
+
return step, report
|
| 233 |
+
return len(mix.LADDER) - 1, report
|
| 234 |
+
|
| 235 |
+
|
| 236 |
+
def build_source(name, size, dedup, quota=None, tag=""):
|
| 237 |
+
"""Fill one source's quota. `tag=".topup"` writes an extra file for fallback top-ups."""
|
| 238 |
+
genre, quotas = mix.MIX[name]
|
| 239 |
+
if quota is None:
|
| 240 |
+
quota = quotas[mix.SIZES[size]] * 1_000_000
|
| 241 |
+
if quota <= 0:
|
| 242 |
+
return None
|
| 243 |
+
out = ROOT / "work" / size / f"{name}{tag}.jsonl"
|
| 244 |
+
done = out.with_suffix(".done")
|
| 245 |
+
if done.exists():
|
| 246 |
+
stats = json.loads(done.read_text())
|
| 247 |
+
if abs(stats["target_M"] - quota / 1e6) < 1e-6:
|
| 248 |
+
for l in out.open(encoding="utf-8"): # reload so global dedup stays consistent
|
| 249 |
+
dedup.seen(json.loads(l)["text"])
|
| 250 |
+
return stats
|
| 251 |
+
base = mix.TIERS[size]
|
| 252 |
+
vocab = set(top_n_list("en", base["vocab"]))
|
| 253 |
+
t0 = time.time()
|
| 254 |
+
step, probe_report = probe(name, size, genre, base, vocab, quota)
|
| 255 |
+
print(f" probe {name}{tag}: start at ladder step {step}; {probe_report[-1]}", flush=True)
|
| 256 |
+
tier, cov = at_step(base, step)
|
| 257 |
+
got = scanned = kept = seen_docs = 0
|
| 258 |
+
openers, rej = Counter(), Counter()
|
| 259 |
+
win_scanned = win_kept = win_docs = 0
|
| 260 |
+
relaxed, abort = [dict(step=step, at_words=0)], None
|
| 261 |
+
with out.open("w", encoding="utf-8") as f:
|
| 262 |
+
for doc in sources.REGISTRY[name](size):
|
| 263 |
+
seen_docs += 1
|
| 264 |
+
win_docs += 1
|
| 265 |
+
w = doc_words(doc)
|
| 266 |
+
scanned += w
|
| 267 |
+
win_scanned += w
|
| 268 |
+
if scanned > quota * mix.MAX_SCAN:
|
| 269 |
+
abort = "scan_cap"
|
| 270 |
+
break
|
| 271 |
+
docs, why = process(doc, genre, tier, vocab, name, cov)
|
| 272 |
+
if why:
|
| 273 |
+
rej[why] += 1
|
| 274 |
+
for text in docs:
|
| 275 |
+
# anti-template: no 4-word opening may exceed 0.5% of the accepted docs
|
| 276 |
+
op = " ".join(text.lower().split()[:4])
|
| 277 |
+
if genre != "poetry" and openers[op] >= max(5, 0.005 * kept):
|
| 278 |
+
rej["opener_cap"] += 1
|
| 279 |
+
continue
|
| 280 |
+
if dedup.seen(text):
|
| 281 |
+
rej["duplicate"] += 1
|
| 282 |
+
continue
|
| 283 |
+
openers[op] += 1
|
| 284 |
+
kept += 1
|
| 285 |
+
n = len(text.split())
|
| 286 |
+
got += n
|
| 287 |
+
win_kept += n
|
| 288 |
+
f.write(json.dumps({"text": text, "genre": genre, "src": name}) + "\n")
|
| 289 |
+
if got >= quota:
|
| 290 |
+
break
|
| 291 |
+
if seen_docs % 2000 == 0:
|
| 292 |
+
print(f" {name}{tag}: {seen_docs} docs, kept {kept}, {got/1e6:.2f}M/{quota/1e6:.2f}M words, "
|
| 293 |
+
f"{time.time()-t0:.0f}s, top rejections {rej.most_common(3)}", flush=True)
|
| 294 |
+
if win_docs >= mix.WINDOW:
|
| 295 |
+
# mechanisms 2+4: if the recent yield cannot finish within the scan budget,
|
| 296 |
+
# climb one ladder step; at the top of the ladder give up on this source
|
| 297 |
+
recent = win_kept / win_scanned if win_scanned else 0
|
| 298 |
+
need = (quota - got) / recent if recent else float("inf")
|
| 299 |
+
if need > quota * mix.MAX_SCAN - scanned:
|
| 300 |
+
if step + 1 < len(mix.LADDER):
|
| 301 |
+
step += 1
|
| 302 |
+
tier, cov = at_step(base, step)
|
| 303 |
+
relaxed.append(dict(step=step, at_words=got))
|
| 304 |
+
print(f" {name}{tag}: yield {100*recent:.1f}% too low, relaxing to step {step}", flush=True)
|
| 305 |
+
else:
|
| 306 |
+
abort = "projected_shortfall"
|
| 307 |
+
break
|
| 308 |
+
win_scanned = win_kept = win_docs = 0
|
| 309 |
+
stats = dict(source=name + tag, genre=genre, target_M=quota / 1e6, words_M=round(got / 1e6, 3),
|
| 310 |
+
docs=kept, scanned_docs=seen_docs, scanned_words_M=round(scanned / 1e6, 1),
|
| 311 |
+
seconds=round(time.time() - t0), stopped="quota" if got >= quota else (abort or "source_exhausted"),
|
| 312 |
+
relaxation=relaxed, probe=probe_report, rejections=dict(rej.most_common()))
|
| 313 |
+
done.write_text(json.dumps(stats))
|
| 314 |
+
return stats
|
| 315 |
+
|
| 316 |
+
|
| 317 |
+
def plan_fallbacks(size, results):
|
| 318 |
+
"""Mechanism 3: spread each source's shortfall over same-genre substitutes, capped."""
|
| 319 |
+
idx = mix.SIZES[size]
|
| 320 |
+
quota = {n: q[idx] * 1e6 for n, (_, q) in mix.MIX.items()}
|
| 321 |
+
target = sum(quota.values())
|
| 322 |
+
got = {r["source"]: r["words_M"] * 1e6 for r in results if "words_M" in r}
|
| 323 |
+
extra, plan, log = Counter(), Counter(), []
|
| 324 |
+
for src, q in quota.items():
|
| 325 |
+
gap = q - got.get(src, 0) if q > 0 else 0
|
| 326 |
+
if gap < 0.01 * q:
|
| 327 |
+
continue
|
| 328 |
+
for sub in mix.FALLBACKS.get(src, []):
|
| 329 |
+
if gap <= 0:
|
| 330 |
+
break
|
| 331 |
+
room = max(mix.EXTRA_FRAC * quota[sub], mix.EXTRA_FLOOR * target) - extra[sub]
|
| 332 |
+
room = min(room, mix.SHARE_CAP * target - got.get(sub, 0) - extra[sub])
|
| 333 |
+
amt = min(gap, room)
|
| 334 |
+
if amt < 0.005 * target:
|
| 335 |
+
continue
|
| 336 |
+
extra[sub] += amt
|
| 337 |
+
plan[sub] += amt
|
| 338 |
+
gap -= amt
|
| 339 |
+
log.append(dict(missing_from=src, filled_by=sub, words_M=round(amt / 1e6, 3)))
|
| 340 |
+
if gap > 0:
|
| 341 |
+
log.append(dict(missing_from=src, unfilled_M=round(gap / 1e6, 3)))
|
| 342 |
+
return plan, log
|
| 343 |
+
|
| 344 |
+
|
| 345 |
+
def main():
|
| 346 |
+
size = sys.argv[1]
|
| 347 |
+
only = sys.argv[2:]
|
| 348 |
+
(ROOT / "work" / size).mkdir(parents=True, exist_ok=True)
|
| 349 |
+
dedup, results = Dedup(), []
|
| 350 |
+
|
| 351 |
+
def run(name, **kw):
|
| 352 |
+
print(f"== {name}{kw.get('tag', '')}", flush=True)
|
| 353 |
+
try:
|
| 354 |
+
r = build_source(name, size, dedup, **kw)
|
| 355 |
+
except Exception as e: # keep going; a failed source is reported, not hidden
|
| 356 |
+
print(f" FAILED {name}: {e!r}", flush=True)
|
| 357 |
+
r = dict(source=name + kw.get("tag", ""), error=repr(e))
|
| 358 |
+
if r:
|
| 359 |
+
print(" ", {k: v for k, v in r.items() if k not in ("probe", "relaxation")}, flush=True)
|
| 360 |
+
results.append(r)
|
| 361 |
+
|
| 362 |
+
for name in mix.MIX:
|
| 363 |
+
if not only or name in only:
|
| 364 |
+
run(name)
|
| 365 |
+
fallback_log = []
|
| 366 |
+
if not only:
|
| 367 |
+
plan, fallback_log = plan_fallbacks(size, [r for r in results if "tag" not in r])
|
| 368 |
+
print("== fallback plan", fallback_log, flush=True)
|
| 369 |
+
for sub, words in plan.items():
|
| 370 |
+
run(sub, quota=words, tag=".topup")
|
| 371 |
+
total = sum(r.get("words_M", 0) for r in results)
|
| 372 |
+
print(f"== total {total:.2f}M words", flush=True)
|
| 373 |
+
(ROOT / "work" / size / "stats.json").write_text(
|
| 374 |
+
json.dumps(dict(total_M=round(total, 3), fallbacks=fallback_log, sources=results), indent=1))
|
| 375 |
+
|
| 376 |
+
|
| 377 |
+
if __name__ == "__main__":
|
| 378 |
+
main()
|
code/mix.py
ADDED
|
@@ -0,0 +1,72 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Mix definition: million words per source for the 10M / 100M / 1000M versions."""
|
| 2 |
+
|
| 3 |
+
SIZES = {"10M": 0, "100M": 1, "1000M": 2}
|
| 4 |
+
|
| 5 |
+
# key: (genre, (10M, 100M, 1000M) quota in million words)
|
| 6 |
+
MIX = {
|
| 7 |
+
"tinystories": ("narrative", (1.5, 8, 40)),
|
| 8 |
+
"simplestories": ("narrative", (2.0, 8, 60)),
|
| 9 |
+
"cosmo_stories": ("narrative", (0, 5, 50)),
|
| 10 |
+
"tinydialogues": ("dialogue", (1.5, 7, 29)),
|
| 11 |
+
"soda": ("dialogue", (1, 9, 75)),
|
| 12 |
+
"ultrachat": ("dialogue", (0, 4, 61)),
|
| 13 |
+
"smoltalk": ("dialogue", (0, 1, 25)),
|
| 14 |
+
"tinyhelen": ("expository", (1.5, 10, 50)),
|
| 15 |
+
"simplewiki": ("expository", (1, 10, 30)),
|
| 16 |
+
"enwiki_intro": ("expository", (0, 0, 50)),
|
| 17 |
+
"cosmo_v2": ("expository", (1, 12, 180)),
|
| 18 |
+
"tiny_textbooks": ("expository", (0, 4, 40)), # uses the synthetic `textbook` column
|
| 19 |
+
"cosmo_v01": ("expository", (0.5, 3, 30)),
|
| 20 |
+
"fineweb_edu": ("expository", (0, 9, 180)),
|
| 21 |
+
"gutenberg_books": ("books", (0, 8, 70)),
|
| 22 |
+
"poetry": ("poetry", (0, 1, 10)),
|
| 23 |
+
"cnn": ("news", (0, 1, 20)),
|
| 24 |
+
}
|
| 25 |
+
|
| 26 |
+
# Complexity tiers per size: reference vocab size, max Flesch-Kincaid grade,
|
| 27 |
+
# document length range (words), max average words per dialogue turn.
|
| 28 |
+
TIERS = {
|
| 29 |
+
"10M": dict(vocab=8000, fk=5, fk_exp=9, doc_max=600, turn=25),
|
| 30 |
+
"100M": dict(vocab=15000, fk=8, fk_exp=10, doc_max=1200, turn=40),
|
| 31 |
+
"1000M": dict(vocab=30000, fk=12, fk_exp=12, doc_max=2000, turn=80),
|
| 32 |
+
}
|
| 33 |
+
# fk_exp: expository text (Cosmopedia textbooks) has median grade ~11 even for school
|
| 34 |
+
# audiences, so fk=5 passed 0.2% of it. Vocabulary coverage still guards simplicity.
|
| 35 |
+
# Assistant chats may average this many times the per-turn word cap of a named dialogue.
|
| 36 |
+
CHAT_TURN_MULT = 4
|
| 37 |
+
|
| 38 |
+
# --- run safety nets (see build.py) ---
|
| 39 |
+
# Relaxation ladder: (extra Flesch-Kincaid grades, change to vocabulary coverage).
|
| 40 |
+
LADDER = [(0, 0.0), (1, -0.02), (2, -0.04), (3, -0.06)]
|
| 41 |
+
PROBE_DOCS = 1000 # docs sampled before a source starts, to pick the starting ladder step
|
| 42 |
+
WINDOW = 5000 # docs per yield check while a source runs
|
| 43 |
+
MAX_SCAN = 40 # hard cap: raw words scanned per source = MAX_SCAN x its quota
|
| 44 |
+
# When a source falls short, same-genre substitutes make up the difference in order.
|
| 45 |
+
# A substitute may add at most max(EXTRA_FRAC x its quota, EXTRA_FLOOR x total target)
|
| 46 |
+
# and never end above SHARE_CAP x total target.
|
| 47 |
+
EXTRA_FRAC, EXTRA_FLOOR, SHARE_CAP = 0.5, 0.05, 0.35
|
| 48 |
+
FALLBACKS = {
|
| 49 |
+
"tinystories": ["simplestories", "cosmo_stories"],
|
| 50 |
+
"simplestories": ["tinystories", "cosmo_stories"],
|
| 51 |
+
"cosmo_stories": ["simplestories", "tinystories"],
|
| 52 |
+
"tinydialogues": ["soda", "smoltalk"],
|
| 53 |
+
"soda": ["tinydialogues", "ultrachat", "smoltalk"],
|
| 54 |
+
"ultrachat": ["smoltalk", "soda"],
|
| 55 |
+
"smoltalk": ["ultrachat", "soda"],
|
| 56 |
+
"tinyhelen": ["tiny_textbooks", "cosmo_v2", "fineweb_edu"],
|
| 57 |
+
"simplewiki": ["tiny_textbooks", "tinyhelen", "cosmo_v2"],
|
| 58 |
+
"enwiki_intro": ["cosmo_v2", "tiny_textbooks", "fineweb_edu"],
|
| 59 |
+
"cosmo_v2": ["tiny_textbooks", "tinyhelen", "fineweb_edu"],
|
| 60 |
+
"cosmo_v01": ["tiny_textbooks", "cosmo_v2", "tinyhelen"],
|
| 61 |
+
"fineweb_edu": ["cosmo_v2", "tiny_textbooks", "tinyhelen"],
|
| 62 |
+
"gutenberg_books": ["tinyhelen", "fineweb_edu"],
|
| 63 |
+
"poetry": [],
|
| 64 |
+
"cnn": ["fineweb_edu", "cosmo_v2"],
|
| 65 |
+
}
|
| 66 |
+
|
| 67 |
+
# min share of (lowercase) tokens inside the reference vocab. The original 95% @ top-5k
|
| 68 |
+
# passed only 3% of SimpleStories and 35% of TinyStories, so it is 90% with larger vocabs.
|
| 69 |
+
VOCAB_COVERAGE = 0.90
|
| 70 |
+
# English-only by construction and made of one-word toddler turns, which fastText misreads
|
| 71 |
+
LID_SKIP = {"tinydialogues"}
|
| 72 |
+
DOC_MIN = 80
|
code/publish.py
ADDED
|
@@ -0,0 +1,257 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Build the Hugging Face release folder (card + parquet + code) and optionally upload it.
|
| 2 |
+
|
| 3 |
+
python publish.py # stage only: out/hf/
|
| 4 |
+
python publish.py --upload # stage and push to a PRIVATE dataset repo (flip to public in the HF UI)
|
| 5 |
+
"""
|
| 6 |
+
import json
|
| 7 |
+
import os
|
| 8 |
+
import shutil
|
| 9 |
+
import sys
|
| 10 |
+
from pathlib import Path
|
| 11 |
+
|
| 12 |
+
ROOT = Path(__file__).resolve().parent.parent
|
| 13 |
+
REPO = "jncarlo/tinymix"
|
| 14 |
+
SIZES = ["1M", "10M", "100M"]
|
| 15 |
+
GENRES = ["narrative", "dialogue", "expository", "books", "poetry", "news"]
|
| 16 |
+
|
| 17 |
+
# source key -> (display name, link, license, role)
|
| 18 |
+
SOURCES = {
|
| 19 |
+
"tinystories": ("TinyStories V2 (GPT-4 split)", "https://huggingface.co/datasets/roneneldan/TinyStories", "CDLA-Sharing-1.0", "child-level stories"),
|
| 20 |
+
"simplestories": ("SimpleStories", "https://huggingface.co/datasets/SimpleStories/SimpleStories", "MIT", "varied simple stories"),
|
| 21 |
+
"cosmo_stories": ("Cosmopedia v0.1, `stories`", "https://huggingface.co/datasets/HuggingFaceTB/cosmopedia", "Apache-2.0", "everyday stories"),
|
| 22 |
+
"tinydialogues": ("TinyDialogues", "https://huggingface.co/datasets/styfeng/TinyDialogues", "MIT", "child-directed dialogue"),
|
| 23 |
+
"soda": ("SODA", "https://huggingface.co/datasets/allenai/soda", "CC-BY-4.0", "social dialogue"),
|
| 24 |
+
"ultrachat": ("UltraChat 200k", "https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k", "MIT", "user/assistant chat"),
|
| 25 |
+
"smoltalk": ("SmolTalk (everyday-conversations, smol-summarize, smol-rewrite)", "https://huggingface.co/datasets/HuggingFaceTB/smoltalk", "Apache-2.0", "chat, summaries, rewrites"),
|
| 26 |
+
"tinyhelen": ("TinyHelen Leaner-Pretrain (book, conversation, textbook, web, wiki)", "https://huggingface.co/datasets/EmpathYang/TinyHelen", "MIT", "simplified real text"),
|
| 27 |
+
"simplewiki": ("Simple English Wikipedia (20231101.simple)", "https://huggingface.co/datasets/wikimedia/wikipedia", "CC-BY-SA-3.0 / GFDL", "encyclopedic, simple"),
|
| 28 |
+
"enwiki_intro": ("English Wikipedia, lead paragraphs", "https://huggingface.co/datasets/wikimedia/wikipedia", "CC-BY-SA-3.0 / GFDL", "encyclopedic"),
|
| 29 |
+
"cosmo_v2": ("Cosmopedia v2 (SmolLM-Corpus)", "https://huggingface.co/datasets/HuggingFaceTB/smollm-corpus", "Apache-2.0 (corpus: ODC-By)", "synthetic textbooks, blogs, stories"),
|
| 30 |
+
"cosmo_v01": ("Cosmopedia v0.1 (wikihow, khanacademy, openstax)", "https://huggingface.co/datasets/HuggingFaceTB/cosmopedia", "Apache-2.0", "procedural and didactic"),
|
| 31 |
+
"tiny_textbooks": ("tiny-textbooks (`textbook` column)", "https://huggingface.co/datasets/nampdn-ai/tiny-textbooks", "Apache-2.0", "short synthetic lessons"),
|
| 32 |
+
"fineweb_edu": ("FineWeb-Edu (sample-10BT, score >= 4)", "https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu", "ODC-By-1.0", "educational web"),
|
| 33 |
+
"gutenberg_books": ("Project Gutenberg (English books)", "https://huggingface.co/datasets/manu/project_gutenberg", "Project Gutenberg License (public domain in the US)", "classic books"),
|
| 34 |
+
"poetry": ("Gutenberg Poetry Corpus", "https://huggingface.co/datasets/biglam/gutenberg-poetry-corpus", "CC0-1.0", "poetry stanzas"),
|
| 35 |
+
"cnn": ("CNN/DailyMail 3.0.0", "https://huggingface.co/datasets/abisee/cnn_dailymail", "Apache-2.0 (articles are (c) CNN / Daily Mail)", "news"),
|
| 36 |
+
}
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
def load(size):
|
| 40 |
+
return json.loads((ROOT / "out" / size / "summary.json").read_text())
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
def tables():
|
| 44 |
+
sums = {s: load(s) for s in SIZES if (ROOT / "out" / s / "summary.json").exists()}
|
| 45 |
+
sizes = list(sums)
|
| 46 |
+
head = "| Subset | Train words | Train docs | ~Tokens (BPE) |\n|---|---|---|---|\n"
|
| 47 |
+
rows = "".join(f"| `{s}` | {v['train_words_M']}M | {v['train_docs']:,} | ~{v['train_words_M']*1.3:.1f}M |\n"
|
| 48 |
+
for s, v in sums.items())
|
| 49 |
+
g = "| Genre | " + " | ".join(sizes) + " |\n|---|" + "---|" * len(sizes) + "\n"
|
| 50 |
+
for gen in GENRES:
|
| 51 |
+
vals = [sums[s]["genre_share_%"].get(gen) for s in sizes]
|
| 52 |
+
if any(vals):
|
| 53 |
+
g += f"| {gen} | " + " | ".join(f"{v or 0}%" for v in vals) + " |\n"
|
| 54 |
+
used = sorted({k for v in sums.values() for k in v["source_words_M"]}, key=lambda k: list(SOURCES).index(k))
|
| 55 |
+
s = "| Source | " + " | ".join(f"{x} (M words)" for x in sizes) + " |\n|---|" + "---|" * len(sizes) + "\n"
|
| 56 |
+
for k in used:
|
| 57 |
+
s += f"| {SOURCES[k][0]} | " + " | ".join(str(sums[x]["source_words_M"].get(k, 0)) for x in sizes) + " |\n"
|
| 58 |
+
lic = "| Source | License | Role |\n|---|---|---|\n" + "".join(
|
| 59 |
+
f"| [{SOURCES[k][0]}]({SOURCES[k][1]}) | {SOURCES[k][2]} | {SOURCES[k][3]} |\n" for k in used)
|
| 60 |
+
return sizes, head + rows, g, s, lic
|
| 61 |
+
|
| 62 |
+
|
| 63 |
+
def results_section():
|
| 64 |
+
runs = ROOT / "runs"
|
| 65 |
+
names = [("10M", "tinymix 10M"), ("base_tinystories_10M", "TinyStories only, 10M words"),
|
| 66 |
+
("base_fineweb_edu_10M", "FineWeb-Edu only, 10M words"),
|
| 67 |
+
("100M", "tinymix 100M"), ("base_fineweb_edu_100M", "FineWeb-Edu only, 100M words")]
|
| 68 |
+
rows = []
|
| 69 |
+
for key, label in names:
|
| 70 |
+
p = runs / key / "results.json"
|
| 71 |
+
if p.exists():
|
| 72 |
+
r = json.loads(p.read_text())
|
| 73 |
+
b = r["valid_bits_per_byte"]
|
| 74 |
+
rows.append(f"| {label} | {r['params_M']}M | {r['tokens_M']}M | " +
|
| 75 |
+
" | ".join(f"{b[g]:.3f}" if g in b else "n/a"
|
| 76 |
+
for g in ("narrative", "dialogue", "expository", "books", "news", "poetry")) + " |\n")
|
| 77 |
+
if not rows:
|
| 78 |
+
return ""
|
| 79 |
+
return ("## Sanity-check training runs\n\n"
|
| 80 |
+
"Small GPTs (nanoGPT-style, own byte-level BPE trained on each training set) trained for the same number of "
|
| 81 |
+
"steps within each size (17M parameters for 10M words, 46M for 100M words), scored in **bits per byte** "
|
| 82 |
+
"(lower is better, comparable across tokenizers) on the held-out validation splits of **this** dataset. "
|
| 83 |
+
"Baselines take the first 10M / 100M words of a single source with only `ftfy` cleaning. They are "
|
| 84 |
+
"out of distribution on the genres they do not cover, so the table shows how much genre coverage matters, "
|
| 85 |
+
"not a leaderboard. One seed, no tuning; the 10M and 100M rows are not comparable with each other.\n\n"
|
| 86 |
+
"| Training data | Params | Train tokens | narrative | dialogue | expository | books | news | poetry |\n|---|---|---|---|---|---|---|---|---|\n"
|
| 87 |
+
+ "".join(rows) + "\n")
|
| 88 |
+
|
| 89 |
+
|
| 90 |
+
def card():
|
| 91 |
+
sizes, size_t, genre_t, src_t, lic_t = tables()
|
| 92 |
+
configs = ""
|
| 93 |
+
for s in sizes:
|
| 94 |
+
configs += f"- config_name: {s}\n data_files:\n - split: train\n path: data/{s}/train.parquet\n"
|
| 95 |
+
for p in sorted((ROOT / "out" / s / "parquet").glob("validation_*.parquet")):
|
| 96 |
+
configs += f" - split: {p.stem}\n path: data/{s}/{p.name}\n"
|
| 97 |
+
return f"""---
|
| 98 |
+
language:
|
| 99 |
+
- en
|
| 100 |
+
license: other
|
| 101 |
+
license_name: mixed-per-source
|
| 102 |
+
license_link: https://huggingface.co/datasets/{REPO}#sources-and-licensing
|
| 103 |
+
pretty_name: tinymix
|
| 104 |
+
size_categories:
|
| 105 |
+
- 100K<n<1M
|
| 106 |
+
task_categories:
|
| 107 |
+
- text-generation
|
| 108 |
+
tags:
|
| 109 |
+
- small-language-models
|
| 110 |
+
- synthetic-data
|
| 111 |
+
- curriculum
|
| 112 |
+
- tinystories
|
| 113 |
+
- babylm
|
| 114 |
+
configs:
|
| 115 |
+
{configs}---
|
| 116 |
+
|
| 117 |
+
# tinymix
|
| 118 |
+
|
| 119 |
+
Small, **genre-balanced English pretraining corpora** in three sizes (about 1M, 10M and 100M words), built for training
|
| 120 |
+
very small language models. The mix combines synthetic child-level stories and dialogue, simplified real text,
|
| 121 |
+
encyclopedic and didactic text, books, poetry and news, all filtered toward clear, self-contained documents. It is
|
| 122 |
+
inspired by TinyStories, SimpleStories, TinyHelen and the BabyLM challenge.
|
| 123 |
+
|
| 124 |
+
What is different from using those datasets directly:
|
| 125 |
+
|
| 126 |
+
- **Several kinds of text in one mix** (narrative, dialogue and chat, expository, books, poetry, news), not only stories or only conversations.
|
| 127 |
+
- **Three sizes built with the same recipe**, with complexity limits that grow with size.
|
| 128 |
+
- **No non-commercial sources.** TinyChat (CC-BY-NC) and the BabyLM corpus (no clear license, includes subtitle and child-speech data) were left out on purpose.
|
| 129 |
+
- **Provenance per document** (`source`, `genre`) and **validation splits per genre**, to see which kind of text a model is weak at.
|
| 130 |
+
- The whole pipeline is included (`code/`) and every filter decision is documented below.
|
| 131 |
+
|
| 132 |
+
## Subsets
|
| 133 |
+
|
| 134 |
+
{size_t}
|
| 135 |
+
Words are whitespace-separated words in `train`. Each subset also has about 0.5% of the documents per genre held out as `validation_<genre>` splits
|
| 136 |
+
(at least one document per genre; the `1M` validation splits are therefore tiny, a few dozen documents, and noisy).
|
| 137 |
+
|
| 138 |
+
```python
|
| 139 |
+
from datasets import load_dataset
|
| 140 |
+
ds = load_dataset("{REPO}", "10M") # splits: train, validation_narrative, validation_dialogue, ...
|
| 141 |
+
print(ds["train"][0]) # {{'text': ..., 'source': 'tinystories', 'genre': 'narrative'}}
|
| 142 |
+
```
|
| 143 |
+
|
| 144 |
+
Columns: `text` (str), `source` (key of the table below), `genre` (`narrative`, `dialogue`, `expository`, `books`, `poetry`, `news`).
|
| 145 |
+
Documents are independent; join them with an end-of-text token when training. Dialogues are formatted one turn per line
|
| 146 |
+
(`Mia: ...`, or `User:` / `Assistant:` for assistant chats).
|
| 147 |
+
|
| 148 |
+
## Composition
|
| 149 |
+
|
| 150 |
+
Share of training words by genre:
|
| 151 |
+
|
| 152 |
+
{genre_t}
|
| 153 |
+
Million words per source:
|
| 154 |
+
|
| 155 |
+
{src_t}
|
| 156 |
+
The 10M and 100M subsets reached their targets (10.0M and 100.0M words including the validation splits). In the 10M build,
|
| 157 |
+
sources that could not fill their quota with the filters (Simple English Wikipedia, Cosmopedia v0.1) were topped up by
|
| 158 |
+
same-genre substitute sources (tiny-textbooks, Cosmopedia v2), capped per source; that is why `tiny-textbooks` appears
|
| 159 |
+
there. The 1M subset keeps the 10M proportions except that Cosmopedia stories (not in the 10M mix) were replaced by TinyStories and SimpleStories.
|
| 160 |
+
|
| 161 |
+
## How it was built
|
| 162 |
+
|
| 163 |
+
1. **Normalisation**: `ftfy`, NFKC, ASCII quotes and dashes, whitespace; URLs, e-mails, phone numbers, code, LaTeX and HTML removed.
|
| 164 |
+
2. **Language**: fastText `lid.176`, English with confidence >= 0.9 (skipped for TinyDialogues, whose one-word toddler turns confuse it).
|
| 165 |
+
3. **Quality heuristics**: Gopher-style rules (mean word length, symbol and digit ratio, repeated lines, repeated bigrams, stop words), minimum length, complete ending (drops truncated generations), a short list of LLM tics (`delve`, `tapestry`, ...) and an explicit-content word list.
|
| 166 |
+
4. **Anti-template**: no 4-word opening may exceed 0.5% of a source's documents.
|
| 167 |
+
5. **Deduplication across all sources**: exact hash and MinHash LSH (5-word shingles, Jaccard >= 0.8).
|
| 168 |
+
6. **Complexity control by size** (vocabulary coverage = share of lowercase tokens inside the top-N English words of `wordfreq`):
|
| 169 |
+
|
| 170 |
+
| | 10M | 100M |
|
| 171 |
+
|---|---|---|
|
| 172 |
+
| Reference vocabulary | top 8k | top 15k |
|
| 173 |
+
| Min vocabulary coverage | 90% | 90% |
|
| 174 |
+
| Max Flesch-Kincaid grade (narrative, news, books) | 5 | 8 |
|
| 175 |
+
| Max Flesch-Kincaid grade (expository) | 9 | 10 |
|
| 176 |
+
| Document length | 80 to 600 words | 80 to 1200 words |
|
| 177 |
+
| Dialogue: mean words per turn | <= 25 | <= 40 |
|
| 178 |
+
|
| 179 |
+
Assistant chats (`User:` / `Assistant:`) may average up to 4 times the per-turn cap and need at least 2 turns (named dialogues need 4),
|
| 180 |
+
because summarisation and rewrite tasks have 2 turns and assistant answers are long.
|
| 181 |
+
|
| 182 |
+
Documents longer than the maximum are split at paragraph borders. The 1M subset is a prefix slice of the 10M build; 10M and 100M were built separately with the same recipe, so they are **not** guaranteed to be subsets of each other (a later fix to the end-of-document check, which wrongly rejected Wikipedia articles that end in a heading, was only applied to sources rebuilt for the 100M run).
|
| 183 |
+
7. **Run safety nets** (used to build the larger subsets without manual tuning): a probe per source picks a starting filter strictness, strictness relaxes in up to three logged steps when yield is too low, and shortfalls go to substitute sources.
|
| 184 |
+
|
| 185 |
+
Thresholds were calibrated on real data: the first values (95% coverage in the top 5k words, Flesch-Kincaid <= 5 for everything)
|
| 186 |
+
accepted only 3% of SimpleStories and 0.6% of Cosmopedia v2, so they were relaxed as described above.
|
| 187 |
+
|
| 188 |
+
## Sources and licensing
|
| 189 |
+
|
| 190 |
+
{lic_t}
|
| 191 |
+
**Licensing and attribution.** This repository is a derivative compilation of the datasets above; each document keeps the
|
| 192 |
+
license of its `source`. Obligations that matter when redistributing or reusing it:
|
| 193 |
+
|
| 194 |
+
- Attribution is required for CC-BY / ODC-By / CDLA sources (listed above).
|
| 195 |
+
- The Wikipedia part (`simplewiki`) is **CC-BY-SA-3.0 / GFDL (share-alike)**. If you need a share-alike-free corpus, filter out `source == "simplewiki"`.
|
| 196 |
+
- Cosmopedia, SmolTalk, UltraChat, TinyStories, SimpleStories, TinyHelen and tiny-textbooks contain **model-generated text**; check the terms of the generating model (GPT-4, GPT-4o-mini, Mixtral and others) for your use case.
|
| 197 |
+
- CNN/DailyMail articles and FineWeb-Edu web pages are third-party content; they are included under the dataset licenses above, which do not transfer rights from the original publishers.
|
| 198 |
+
- Project Gutenberg texts had the Project Gutenberg header and footer removed; they are public domain in the United States, check your jurisdiction.
|
| 199 |
+
|
| 200 |
+
No personal data was added; e-mails, phone numbers and URLs were removed, but residual personal information in web-derived sources cannot be ruled out.
|
| 201 |
+
|
| 202 |
+
{results_section()}## Limitations
|
| 203 |
+
|
| 204 |
+
- **Not done**: perplexity filtering with an n-gram LM, LLM-judge coherence scoring, semantic deduplication, cluster caps and benchmark decontamination. Expect some residual near-duplicates and some low-quality documents.
|
| 205 |
+
- Some `tinyhelen` chunks start mid-thought, because the simplified web and book documents are chunks of longer texts.
|
| 206 |
+
- Synthetic sources share model-specific style and topics (names, morals, classroom settings). The anti-template cap reduces this but does not remove it.
|
| 207 |
+
- The toxicity filter is a short word list, not a classifier. Chat and web sources can still contain mildly sensitive content.
|
| 208 |
+
- English only, mostly simple modern language; it is a poor fit for models meant to handle technical, legal or archaic text.
|
| 209 |
+
- Word counts are whitespace words; token counts depend on your tokenizer.
|
| 210 |
+
|
| 211 |
+
## Reproduce
|
| 212 |
+
|
| 213 |
+
`code/` contains the build scripts (`mix.py` quotas and tiers, `sources.py` adapters, `build.py` pipeline, `assemble.py`,
|
| 214 |
+
`publish.py`, `train.py` for the sanity-check runs) and `requirements.txt`. Building a subset: `python build.py 10M`,
|
| 215 |
+
then `python assemble.py 10M`. The build streams the source datasets from the Hub.
|
| 216 |
+
|
| 217 |
+
## Acknowledgements
|
| 218 |
+
|
| 219 |
+
Built on the work behind TinyStories, SimpleStories, TinyDialogues, SODA, UltraChat, SmolTalk and SmolLM-Corpus
|
| 220 |
+
(Cosmopedia, FineWeb-Edu), TinyHelen, tiny-textbooks, Project Gutenberg, the Gutenberg Poetry Corpus, Wikipedia and
|
| 221 |
+
CNN/DailyMail. Please cite the original datasets when you use this one.
|
| 222 |
+
"""
|
| 223 |
+
|
| 224 |
+
|
| 225 |
+
def stage():
|
| 226 |
+
out = ROOT / "out" / "hf"
|
| 227 |
+
if out.exists():
|
| 228 |
+
shutil.rmtree(out)
|
| 229 |
+
(out / "code").mkdir(parents=True)
|
| 230 |
+
sizes = [s for s in SIZES if (ROOT / "out" / s / "parquet" / "train.parquet").exists()]
|
| 231 |
+
for s in sizes:
|
| 232 |
+
shutil.copytree(ROOT / "out" / s / "parquet", out / "data" / s)
|
| 233 |
+
for f in ("mix.py", "sources.py", "build.py", "assemble.py", "baselines.py", "train.py", "publish.py"):
|
| 234 |
+
shutil.copy(ROOT / "scripts" / f, out / "code" / f)
|
| 235 |
+
shutil.copy(ROOT / "requirements.txt", out / "code" / "requirements.txt")
|
| 236 |
+
(out / "README.md").write_text(card(), encoding="utf-8")
|
| 237 |
+
print("staged", sizes, "->", out)
|
| 238 |
+
return out
|
| 239 |
+
|
| 240 |
+
|
| 241 |
+
def main():
|
| 242 |
+
out = stage()
|
| 243 |
+
if "--upload" in sys.argv:
|
| 244 |
+
for l in (ROOT / ".env").read_text(encoding="utf-8").splitlines():
|
| 245 |
+
if "=" in l and not l.lstrip().startswith("#"):
|
| 246 |
+
k, v = l.split("=", 1)
|
| 247 |
+
os.environ.setdefault(k.strip(), v.strip().strip("'\""))
|
| 248 |
+
from huggingface_hub import HfApi
|
| 249 |
+
api = HfApi(token=os.environ["HF_TOKEN"])
|
| 250 |
+
api.create_repo(REPO, repo_type="dataset", private=True, exist_ok=True)
|
| 251 |
+
api.upload_folder(folder_path=str(out), repo_id=REPO, repo_type="dataset",
|
| 252 |
+
commit_message="tinymix: " + ", ".join(p.name for p in (out / "data").iterdir()))
|
| 253 |
+
print(f"uploaded to https://huggingface.co/datasets/{REPO} (private)")
|
| 254 |
+
|
| 255 |
+
|
| 256 |
+
if __name__ == "__main__":
|
| 257 |
+
main()
|
code/requirements.txt
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
datasets==5.1.0
|
| 2 |
+
huggingface_hub>=1.5,<2
|
| 3 |
+
tokenizers==0.22.2
|
| 4 |
+
ftfy==6.3.1
|
| 5 |
+
datasketch==2.0.0
|
| 6 |
+
wordfreq==3.1.1
|
| 7 |
+
pyarrow
|
| 8 |
+
numpy==2.1.3
|
| 9 |
+
fasttext-wheel
|
| 10 |
+
vastai
|
code/sources.py
ADDED
|
@@ -0,0 +1,193 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Streaming adapters. Each yields dicts: {"text"} or {"turns": [(speaker, text)]}, optional "title"."""
|
| 2 |
+
import ast
|
| 3 |
+
import random
|
| 4 |
+
import re
|
| 5 |
+
import zlib
|
| 6 |
+
|
| 7 |
+
from datasets import load_dataset
|
| 8 |
+
from huggingface_hub import HfFileSystem
|
| 9 |
+
|
| 10 |
+
SEED = 0
|
| 11 |
+
EOT = "<|endoftext|>"
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
def _ds(*args, shuffle=False, buffer=20000, **kw):
|
| 15 |
+
ds = load_dataset(*args, streaming=True, **kw)
|
| 16 |
+
return ds.shuffle(seed=SEED, buffer_size=buffer) if shuffle else ds
|
| 17 |
+
|
| 18 |
+
|
| 19 |
+
def _lines(path):
|
| 20 |
+
with HfFileSystem().open(path, "r", encoding="utf-8", errors="replace") as f:
|
| 21 |
+
yield from f
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
def tinystories(size):
|
| 25 |
+
buf = []
|
| 26 |
+
for line in _lines("datasets/roneneldan/TinyStories/TinyStoriesV2-GPT4-train.txt"):
|
| 27 |
+
if EOT in line:
|
| 28 |
+
text = " ".join(buf).strip()
|
| 29 |
+
buf = []
|
| 30 |
+
if text:
|
| 31 |
+
# the source lost curly quotes as U+FFFD: between letters it was an apostrophe
|
| 32 |
+
text = re.sub(r"(?<=\w)�(?=\w)", "'", text).replace("�", '"')
|
| 33 |
+
yield {"text": text}
|
| 34 |
+
else:
|
| 35 |
+
buf.append(line.strip())
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
def simplestories(size):
|
| 39 |
+
for r in _ds("SimpleStories/SimpleStories", split="train", shuffle=True):
|
| 40 |
+
yield {"text": r["story"]}
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
def cosmo_stories(size):
|
| 44 |
+
for r in _ds("HuggingFaceTB/cosmopedia", "stories", split="train", shuffle=True):
|
| 45 |
+
yield {"text": r["text"], "audience": r.get("audience")}
|
| 46 |
+
|
| 47 |
+
|
| 48 |
+
def tinydialogues(size):
|
| 49 |
+
for line in _lines("datasets/styfeng/TinyDialogues/tinydialogue_train_ordered.txt"):
|
| 50 |
+
for doc in line.split(EOT):
|
| 51 |
+
turns = []
|
| 52 |
+
for t in doc.split("\\n\\n"):
|
| 53 |
+
m = re.match(r'\s*\*\*(.+?)\*\*\s*:\s*(.+)', t.strip(), re.S)
|
| 54 |
+
if m:
|
| 55 |
+
turns.append((m.group(1).strip(), m.group(2).strip().strip('"')))
|
| 56 |
+
if turns:
|
| 57 |
+
yield {"turns": turns}
|
| 58 |
+
|
| 59 |
+
|
| 60 |
+
def soda(size):
|
| 61 |
+
for r in _ds("allenai/soda", split="train", shuffle=True):
|
| 62 |
+
d, s = r["dialogue"], r["speakers"]
|
| 63 |
+
d = ast.literal_eval(d) if isinstance(d, str) else d
|
| 64 |
+
s = ast.literal_eval(s) if isinstance(s, str) else s
|
| 65 |
+
if d and len(d) == len(s) and 4 <= len(d) <= 12:
|
| 66 |
+
yield {"turns": list(zip(s, d))}
|
| 67 |
+
|
| 68 |
+
|
| 69 |
+
def _chat(msgs):
|
| 70 |
+
names = {"user": "User", "assistant": "Assistant"}
|
| 71 |
+
return [(names[m["role"]], m["content"]) for m in msgs if m["role"] in names]
|
| 72 |
+
|
| 73 |
+
|
| 74 |
+
def ultrachat(size):
|
| 75 |
+
for r in _ds("HuggingFaceH4/ultrachat_200k", split="train_sft", shuffle=True):
|
| 76 |
+
yield {"turns": _chat(r["messages"])}
|
| 77 |
+
|
| 78 |
+
|
| 79 |
+
def smoltalk(size):
|
| 80 |
+
# interleave the three configs so none is exhausted first
|
| 81 |
+
its = [iter(_ds("HuggingFaceTB/smoltalk", c, split="train", shuffle=True))
|
| 82 |
+
for c in ("everyday-conversations", "smol-summarize", "smol-rewrite")]
|
| 83 |
+
while its:
|
| 84 |
+
for it in list(its):
|
| 85 |
+
try:
|
| 86 |
+
yield {"turns": _chat(next(it)["messages"])}
|
| 87 |
+
except StopIteration:
|
| 88 |
+
its.remove(it)
|
| 89 |
+
|
| 90 |
+
|
| 91 |
+
def tinyhelen(size):
|
| 92 |
+
base = "hf://datasets/EmpathYang/TinyHelen/data/leaner/100M/train/"
|
| 93 |
+
fs = HfFileSystem()
|
| 94 |
+
files = sorted(p for p in fs.ls("datasets/EmpathYang/TinyHelen/data/leaner/100M/train", detail=False)
|
| 95 |
+
if any(k in p.split("/")[-1] for k in ("book", "conversation", "textbook", "web", "wiki")))
|
| 96 |
+
streams = [iter(load_dataset("json", data_files=base + p.split("/")[-1], split="train", streaming=True))
|
| 97 |
+
for p in files]
|
| 98 |
+
# round-robin over the genre files so the mix is balanced
|
| 99 |
+
while streams:
|
| 100 |
+
for it in list(streams):
|
| 101 |
+
try:
|
| 102 |
+
yield {"text": next(it)["text"]}
|
| 103 |
+
except StopIteration:
|
| 104 |
+
streams.remove(it)
|
| 105 |
+
|
| 106 |
+
|
| 107 |
+
def simplewiki(size):
|
| 108 |
+
for r in _ds("wikimedia/wikipedia", "20231101.simple", split="train", shuffle=True):
|
| 109 |
+
yield {"text": r["text"], "title": r["title"]}
|
| 110 |
+
|
| 111 |
+
|
| 112 |
+
def enwiki_intro(size):
|
| 113 |
+
for r in _ds("wikimedia/wikipedia", "20231101.en", split="train", shuffle=True):
|
| 114 |
+
intro = r["text"].split("\n\n")[0]
|
| 115 |
+
yield {"text": intro, "title": r["title"]}
|
| 116 |
+
|
| 117 |
+
|
| 118 |
+
def cosmo_v2(size):
|
| 119 |
+
kids = {"young_children", "middle_school_students", "grade_school_students", "children"}
|
| 120 |
+
for r in _ds("HuggingFaceTB/smollm-corpus", "cosmopedia-v2", split="train", shuffle=True):
|
| 121 |
+
if size == "10M" and r["audience"] not in kids:
|
| 122 |
+
continue
|
| 123 |
+
yield {"text": r["text"], "audience": r["audience"], "format": r["format"]}
|
| 124 |
+
|
| 125 |
+
|
| 126 |
+
def cosmo_v01(size):
|
| 127 |
+
its = [iter(_ds("HuggingFaceTB/cosmopedia", c, split="train", shuffle=True))
|
| 128 |
+
for c in ("wikihow", "khanacademy", "openstax")]
|
| 129 |
+
while its:
|
| 130 |
+
for it in list(its):
|
| 131 |
+
try:
|
| 132 |
+
yield {"text": next(it)["text"]}
|
| 133 |
+
except StopIteration:
|
| 134 |
+
its.remove(it)
|
| 135 |
+
|
| 136 |
+
|
| 137 |
+
def tiny_textbooks(size):
|
| 138 |
+
# `text` is raw web (RefinedWeb); `textbook` is the synthetic lesson we want
|
| 139 |
+
for r in _ds("nampdn-ai/tiny-textbooks", split="train", shuffle=True):
|
| 140 |
+
yield {"text": r["textbook"]}
|
| 141 |
+
|
| 142 |
+
|
| 143 |
+
def fineweb_edu(size):
|
| 144 |
+
for r in _ds("HuggingFaceFW/fineweb-edu", "sample-10BT", split="train"):
|
| 145 |
+
if r["int_score"] >= 4:
|
| 146 |
+
yield {"text": r["text"]}
|
| 147 |
+
|
| 148 |
+
|
| 149 |
+
def gutenberg_books(size):
|
| 150 |
+
# whole books per row: a 20k-row shuffle buffer overflows Arrow's 2 GB string offsets
|
| 151 |
+
for r in _ds("manu/project_gutenberg", split="en", shuffle=True, buffer=8):
|
| 152 |
+
t = r["text"]
|
| 153 |
+
a = re.search(r"\*\*\* ?START OF (THE|THIS) PROJECT GUTENBERG[^\n]*\n", t)
|
| 154 |
+
b = re.search(r"\*\*\* ?END OF (THE|THIS) PROJECT GUTENBERG", t)
|
| 155 |
+
if a:
|
| 156 |
+
t = t[a.end():]
|
| 157 |
+
if b:
|
| 158 |
+
t = t[:t.rfind("*** END")] if "*** END" in t else t[:b.start()]
|
| 159 |
+
paras = [re.sub(r"\s*\n\s*", " ", p).strip() for p in re.split(r"\n\s*\n", t)]
|
| 160 |
+
paras = [p for p in paras if p and not (p.isupper() and len(p) < 80)]
|
| 161 |
+
chunk, n = [], 0
|
| 162 |
+
for p in paras:
|
| 163 |
+
chunk.append(p)
|
| 164 |
+
n += len(p.split())
|
| 165 |
+
if n >= 300:
|
| 166 |
+
yield {"text": "\n\n".join(chunk)}
|
| 167 |
+
chunk, n = [], 0
|
| 168 |
+
|
| 169 |
+
|
| 170 |
+
def poetry(size):
|
| 171 |
+
rng = random.Random(SEED)
|
| 172 |
+
cur, lines, target = None, [], rng.randint(4, 20)
|
| 173 |
+
for r in _ds("biglam/gutenberg-poetry-corpus", split="train"):
|
| 174 |
+
if r["gutenberg_id"] != cur:
|
| 175 |
+
cur, lines = r["gutenberg_id"], []
|
| 176 |
+
lines.append(r["line"].strip())
|
| 177 |
+
if len(lines) >= target:
|
| 178 |
+
yield {"text": "\n".join(lines)}
|
| 179 |
+
lines, target = [], rng.randint(4, 20)
|
| 180 |
+
|
| 181 |
+
|
| 182 |
+
def cnn(size):
|
| 183 |
+
for r in _ds("abisee/cnn_dailymail", "3.0.0", split="train", shuffle=True):
|
| 184 |
+
text = r["article"]
|
| 185 |
+
if zlib.crc32(r["id"].encode()) % 10 < 3:
|
| 186 |
+
text += "\n\nIn short: " + " ".join(r["highlights"].split("\n"))
|
| 187 |
+
yield {"text": text}
|
| 188 |
+
|
| 189 |
+
|
| 190 |
+
REGISTRY = {f.__name__: f for f in (
|
| 191 |
+
tinystories, simplestories, cosmo_stories, tinydialogues, soda, ultrachat, smoltalk,
|
| 192 |
+
tinyhelen, simplewiki, enwiki_intro, cosmo_v2, cosmo_v01, tiny_textbooks, fineweb_edu,
|
| 193 |
+
gutenberg_books, poetry, cnn)}
|
code/train.py
ADDED
|
@@ -0,0 +1,180 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Tiny GPT (nanoGPT-style) on one dataset size, with its own BPE.
|
| 2 |
+
|
| 3 |
+
python train.py 1M [--vocab 4096 --layers 6 --heads 6 --dim 192 --epochs 10]
|
| 4 |
+
|
| 5 |
+
Reports validation loss per genre as bits per byte, which is comparable across
|
| 6 |
+
datasets even though each one has its own tokenizer.
|
| 7 |
+
"""
|
| 8 |
+
import argparse
|
| 9 |
+
import json
|
| 10 |
+
import math
|
| 11 |
+
import time
|
| 12 |
+
from pathlib import Path
|
| 13 |
+
|
| 14 |
+
import torch
|
| 15 |
+
import torch.nn as nn
|
| 16 |
+
import torch.nn.functional as F
|
| 17 |
+
from tokenizers import Tokenizer, models, pre_tokenizers, decoders, trainers
|
| 18 |
+
|
| 19 |
+
ROOT = Path(__file__).resolve().parent.parent
|
| 20 |
+
EOT = "<|endoftext|>"
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
def get_tokenizer(data, vocab, path):
|
| 24 |
+
if path.exists():
|
| 25 |
+
return Tokenizer.from_file(str(path))
|
| 26 |
+
tok = Tokenizer(models.BPE())
|
| 27 |
+
tok.pre_tokenizer = pre_tokenizers.ByteLevel(add_prefix_space=False)
|
| 28 |
+
tok.decoder = decoders.ByteLevel()
|
| 29 |
+
tr = trainers.BpeTrainer(vocab_size=vocab, special_tokens=[EOT],
|
| 30 |
+
initial_alphabet=pre_tokenizers.ByteLevel.alphabet())
|
| 31 |
+
tok.train_from_iterator((d for d in data.split(EOT) if d.strip()), tr)
|
| 32 |
+
tok.save(str(path))
|
| 33 |
+
return tok
|
| 34 |
+
|
| 35 |
+
|
| 36 |
+
def encode(tok, text):
|
| 37 |
+
ids = []
|
| 38 |
+
eot = tok.token_to_id(EOT)
|
| 39 |
+
for doc in text.split(EOT):
|
| 40 |
+
if doc.strip():
|
| 41 |
+
ids += tok.encode(doc.strip("\n")).ids + [eot]
|
| 42 |
+
return torch.tensor(ids, dtype=torch.long)
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
class Block(nn.Module):
|
| 46 |
+
def __init__(self, d, h):
|
| 47 |
+
super().__init__()
|
| 48 |
+
self.h = h
|
| 49 |
+
self.ln1, self.ln2 = nn.LayerNorm(d), nn.LayerNorm(d)
|
| 50 |
+
self.qkv, self.proj = nn.Linear(d, 3 * d, bias=False), nn.Linear(d, d, bias=False)
|
| 51 |
+
self.fc, self.out = nn.Linear(d, 4 * d, bias=False), nn.Linear(4 * d, d, bias=False)
|
| 52 |
+
|
| 53 |
+
def forward(self, x):
|
| 54 |
+
B, T, C = x.shape
|
| 55 |
+
q, k, v = self.qkv(self.ln1(x)).view(B, T, 3, self.h, C // self.h).permute(2, 0, 3, 1, 4)
|
| 56 |
+
x = x + self.proj(F.scaled_dot_product_attention(q, k, v, is_causal=True)
|
| 57 |
+
.transpose(1, 2).reshape(B, T, C))
|
| 58 |
+
return x + self.out(F.gelu(self.fc(self.ln2(x))))
|
| 59 |
+
|
| 60 |
+
|
| 61 |
+
class GPT(nn.Module):
|
| 62 |
+
def __init__(self, vocab, block, layers, heads, dim):
|
| 63 |
+
super().__init__()
|
| 64 |
+
self.block = block
|
| 65 |
+
self.tok, self.pos = nn.Embedding(vocab, dim), nn.Embedding(block, dim)
|
| 66 |
+
self.blocks = nn.ModuleList(Block(dim, heads) for _ in range(layers))
|
| 67 |
+
self.ln = nn.LayerNorm(dim)
|
| 68 |
+
self.head = nn.Linear(dim, vocab, bias=False)
|
| 69 |
+
self.head.weight = self.tok.weight # tied embeddings
|
| 70 |
+
self.apply(self._init)
|
| 71 |
+
|
| 72 |
+
@staticmethod
|
| 73 |
+
def _init(m):
|
| 74 |
+
if isinstance(m, (nn.Linear, nn.Embedding)):
|
| 75 |
+
nn.init.normal_(m.weight, std=0.02)
|
| 76 |
+
|
| 77 |
+
def forward(self, idx, targets=None):
|
| 78 |
+
x = self.tok(idx) + self.pos(torch.arange(idx.size(1), device=idx.device))
|
| 79 |
+
for b in self.blocks:
|
| 80 |
+
x = b(x)
|
| 81 |
+
logits = self.head(self.ln(x))
|
| 82 |
+
loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1)) if targets is not None else None
|
| 83 |
+
return logits, loss
|
| 84 |
+
|
| 85 |
+
@torch.no_grad()
|
| 86 |
+
def generate(self, idx, n, temp=0.8, top_k=40):
|
| 87 |
+
for _ in range(n):
|
| 88 |
+
logits, _ = self(idx[:, -self.block:])
|
| 89 |
+
logits = logits[:, -1] / temp
|
| 90 |
+
v, _ = torch.topk(logits, top_k)
|
| 91 |
+
logits[logits < v[:, [-1]]] = -float("inf")
|
| 92 |
+
idx = torch.cat([idx, torch.multinomial(F.softmax(logits, -1), 1)], 1)
|
| 93 |
+
return idx
|
| 94 |
+
|
| 95 |
+
|
| 96 |
+
@torch.no_grad()
|
| 97 |
+
def eval_bpb(model, tok, text, block, dev, amp):
|
| 98 |
+
"""Bits per byte over a validation file (non-overlapping windows)."""
|
| 99 |
+
ids = encode(tok, text)
|
| 100 |
+
nbytes = len(text.replace(EOT, "").encode("utf-8"))
|
| 101 |
+
model.eval()
|
| 102 |
+
total, count = 0.0, 0
|
| 103 |
+
for i in range(0, len(ids) - 1, block):
|
| 104 |
+
x = ids[i:i + block].to(dev)[None]
|
| 105 |
+
y = ids[i + 1:i + block + 1].to(dev)[None]
|
| 106 |
+
x = x[:, :y.size(1)]
|
| 107 |
+
with torch.autocast("cuda", dtype=torch.bfloat16, enabled=amp):
|
| 108 |
+
_, loss = model(x, y)
|
| 109 |
+
total += loss.item() * y.numel()
|
| 110 |
+
count += y.numel()
|
| 111 |
+
model.train()
|
| 112 |
+
# summed nats over all tokens, divided by bytes: independent of the tokenizer
|
| 113 |
+
return total / math.log(2) / nbytes
|
| 114 |
+
|
| 115 |
+
|
| 116 |
+
def main():
|
| 117 |
+
ap = argparse.ArgumentParser()
|
| 118 |
+
ap.add_argument("size")
|
| 119 |
+
ap.add_argument("--vocab", type=int, default=4096)
|
| 120 |
+
ap.add_argument("--layers", type=int, default=6)
|
| 121 |
+
ap.add_argument("--heads", type=int, default=6)
|
| 122 |
+
ap.add_argument("--dim", type=int, default=192)
|
| 123 |
+
ap.add_argument("--block", type=int, default=256)
|
| 124 |
+
ap.add_argument("--batch", type=int, default=64)
|
| 125 |
+
ap.add_argument("--epochs", type=float, default=10)
|
| 126 |
+
ap.add_argument("--steps", type=int, default=0, help="fixed step count (overrides --epochs); used to match compute across datasets")
|
| 127 |
+
ap.add_argument("--lr", type=float, default=1e-3)
|
| 128 |
+
ap.add_argument("--tag", default="")
|
| 129 |
+
ap.add_argument("--data", default=None, help="folder with train.txt (default out/<size>)")
|
| 130 |
+
ap.add_argument("--valid", default=None, help="folder with valid_*.txt (default: --data); "
|
| 131 |
+
"baselines are scored on OUR validation sets")
|
| 132 |
+
a = ap.parse_args()
|
| 133 |
+
|
| 134 |
+
torch.manual_seed(0)
|
| 135 |
+
dev = "cuda" if torch.cuda.is_available() else "cpu"
|
| 136 |
+
amp = dev == "cuda" and torch.cuda.is_bf16_supported()
|
| 137 |
+
data_dir = Path(a.data) if a.data else ROOT / "out" / a.size
|
| 138 |
+
run = ROOT / "runs" / f"{a.size}{a.tag}"
|
| 139 |
+
run.mkdir(parents=True, exist_ok=True)
|
| 140 |
+
|
| 141 |
+
train_text = (data_dir / "train.txt").read_text(encoding="utf-8")
|
| 142 |
+
tok = get_tokenizer(train_text, a.vocab, run / "tokenizer.json")
|
| 143 |
+
train = encode(tok, train_text)
|
| 144 |
+
valid_dir = Path(a.valid) if a.valid else data_dir
|
| 145 |
+
valids = {p.stem.replace("valid_", ""): p.read_text(encoding="utf-8") for p in sorted(valid_dir.glob("valid_*.txt"))}
|
| 146 |
+
model = GPT(tok.get_vocab_size(), a.block, a.layers, a.heads, a.dim).to(dev)
|
| 147 |
+
n_params = sum(p.numel() for p in model.parameters())
|
| 148 |
+
steps = a.steps or int(a.epochs * len(train) / (a.batch * a.block))
|
| 149 |
+
print(f"{a.size}: {len(train)/1e6:.2f}M tokens, {n_params/1e6:.2f}M params, {steps} steps, device {dev}, bf16 {amp}", flush=True)
|
| 150 |
+
|
| 151 |
+
opt = torch.optim.AdamW(model.parameters(), lr=a.lr, betas=(0.9, 0.95), weight_decay=0.1)
|
| 152 |
+
sched = lambda s: a.lr * min(1, (s + 1) / 100) * (0.1 + 0.9 * 0.5 * (1 + math.cos(math.pi * s / steps)))
|
| 153 |
+
t0 = time.time()
|
| 154 |
+
for s in range(steps):
|
| 155 |
+
for g in opt.param_groups:
|
| 156 |
+
g["lr"] = sched(s)
|
| 157 |
+
ix = torch.randint(len(train) - a.block - 1, (a.batch,))
|
| 158 |
+
x = torch.stack([train[i:i + a.block] for i in ix]).to(dev)
|
| 159 |
+
y = torch.stack([train[i + 1:i + a.block + 1] for i in ix]).to(dev)
|
| 160 |
+
with torch.autocast("cuda", dtype=torch.bfloat16, enabled=amp):
|
| 161 |
+
_, loss = model(x, y)
|
| 162 |
+
opt.zero_grad(set_to_none=True)
|
| 163 |
+
loss.backward()
|
| 164 |
+
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
|
| 165 |
+
opt.step()
|
| 166 |
+
if s % 100 == 0 or s == steps - 1:
|
| 167 |
+
print(f"step {s}/{steps} loss {loss.item():.3f} {time.time()-t0:.0f}s", flush=True)
|
| 168 |
+
|
| 169 |
+
res = {"size": a.size, "params_M": round(n_params / 1e6, 2), "tokens_M": round(len(train) / 1e6, 2),
|
| 170 |
+
"steps": steps, "seconds": round(time.time() - t0),
|
| 171 |
+
"valid_bits_per_byte": {g: round(eval_bpb(model, tok, t, a.block, dev, amp), 4) for g, t in valids.items()}}
|
| 172 |
+
prompt = torch.tensor([[tok.token_to_id(EOT)]], device=dev)
|
| 173 |
+
res["sample"] = tok.decode(model.generate(prompt, 150)[0].tolist())
|
| 174 |
+
(run / "results.json").write_text(json.dumps(res, indent=1), encoding="utf-8")
|
| 175 |
+
torch.save(model.state_dict(), run / "model.pt")
|
| 176 |
+
print(json.dumps(res, indent=1))
|
| 177 |
+
|
| 178 |
+
|
| 179 |
+
if __name__ == "__main__":
|
| 180 |
+
main()
|
data/100M/train.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:4a2a6e05e47b4109ff6dcf8532ffeb14fbe9481518954ba1f894549048301289
|
| 3 |
+
size 241518580
|
data/100M/validation_books.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:17ea604e9a0673fb38a570edc5919a2e7112b65ddd1ff51ee75ad49eb05cd894
|
| 3 |
+
size 103933
|
data/100M/validation_dialogue.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:bade9587417964ad14caf9538eafc53c425d4889b6c40aa4ad05efe3e7588831
|
| 3 |
+
size 216293
|
data/100M/validation_expository.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:70ab0434691b010f2140d4792843975ba6065f617b6b661c2224ce0082cc35fa
|
| 3 |
+
size 596237
|
data/100M/validation_narrative.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a87e9ce6126904af3d3ff4faf50ebd43e8bc53835164df06461c639678a147bb
|
| 3 |
+
size 232296
|
data/100M/validation_news.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:844e4285c5fb536acbb6383a5614f4a7b22295cd7726bc907b8a4bbe5c1aea00
|
| 3 |
+
size 20266
|
data/100M/validation_poetry.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3d8605adec8b235dcf2c8d14161073b1ff270d1ba63669b3fea4b26b6358f4f2
|
| 3 |
+
size 16992
|
data/10M/train.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d2288906074861a0d74a3240d6e7493838979cdbb7e0d1d38a7c9c39a749e952
|
| 3 |
+
size 22521290
|
data/10M/validation_dialogue.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:e4ed00f76d5dd7ab7dcdad239414a91e5eb2615c61d09696368e1265f71b3194
|
| 3 |
+
size 30704
|
data/10M/validation_expository.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1aa892bbbddd49e23abbb66c6f5b142de188f4e1feb56fb1722aa89d72dc8dec
|
| 3 |
+
size 52271
|
data/10M/validation_narrative.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d0e07ef77350053be742b174c8c14986775afe7321c885ad8f654c9107d05fad
|
| 3 |
+
size 43973
|
data/1M/train.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a128c794cebfe7de2cd6e1feb50e798b0fbea0db79e527c960634699ec66de95
|
| 3 |
+
size 2293603
|
data/1M/validation_dialogue.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:4a07005ab0813d76f3f649d3e9f118ddd6fc500631f8bc2c6ace815cac0d3b71
|
| 3 |
+
size 7082
|
data/1M/validation_expository.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:893a340e5d1271551e347b2265787b8849a7888011a26ae5296285e34f8488d0
|
| 3 |
+
size 12131
|
data/1M/validation_narrative.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a1ec4dffab1b3e2117c8c06987ec396f50d5dc0392bfeed6fcc543705fb3cb2e
|
| 3 |
+
size 7624
|