Consolidated 1B tokenizer program card (vocab profiles, layout, siblings, license)
Browse files
README.md
CHANGED
|
@@ -1,309 +1,366 @@
|
|
| 1 |
-
|
| 2 |
-
|
| 3 |
-
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
```
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
``
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
``
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
``
|
| 42 |
-
|
| 43 |
-
``
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
```
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
#
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
#
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
|
|
| 124 |
-
|
|
| 125 |
-
|
|
| 126 |
-
|
|
| 127 |
-
|
| 128 |
-
|
| 129 |
-
|
| 130 |
-
|
| 131 |
-
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
|
| 135 |
-
|
| 136 |
-
|
|
| 137 |
-
|
|
| 138 |
-
|
|
| 139 |
-
|
|
| 140 |
-
|
| 141 |
-
|
| 142 |
-
|
| 143 |
-
|
| 144 |
-
|
| 145 |
-
|
| 146 |
-
|
| 147 |
-
|
| 148 |
-
|
| 149 |
-
|
| 150 |
-
|
| 151 |
-
|
| 152 |
-
|
| 153 |
-
|
| 154 |
-
|
| 155 |
-
|
| 156 |
-
|
| 157 |
-
|
| 158 |
-
|
| 159 |
-
|
| 160 |
-
|
| 161 |
-
|
| 162 |
-
|
| 163 |
-
|
| 164 |
-
|
| 165 |
-
|
| 166 |
-
|
| 167 |
-
|
| 168 |
-
|
| 169 |
-
|
| 170 |
-
|
| 171 |
-
|
| 172 |
-
|
| 173 |
-
|
| 174 |
-
|
| 175 |
-
|
| 176 |
-
|
| 177 |
-
|
| 178 |
-
|
| 179 |
-
|
| 180 |
-
|
| 181 |
-
|
| 182 |
-
|
| 183 |
-
|
| 184 |
-
|
| 185 |
-
|
| 186 |
-
|
| 187 |
-
|
| 188 |
-
|
| 189 |
-
|
| 190 |
-
|
| 191 |
-
|
| 192 |
-
|
| 193 |
-
|
| 194 |
-
|
| 195 |
-
|
| 196 |
-
|
| 197 |
-
|
| 198 |
-
|
| 199 |
-
|
| 200 |
-
|
| 201 |
-
* **
|
| 202 |
-
* **
|
| 203 |
-
* **
|
| 204 |
-
|
| 205 |
-
|
| 206 |
-
|
| 207 |
-
|
| 208 |
-
|
| 209 |
-
|
| 210 |
-
|
| 211 |
-
|
| 212 |
-
|
| 213 |
-
|
| 214 |
-
|
| 215 |
-
|
| 216 |
-
|
| 217 |
-
|
| 218 |
-
|
| 219 |
-
|
| 220 |
-
|
| 221 |
-
|
| 222 |
-
|
| 223 |
-
|
| 224 |
-
|
| 225 |
-
|
| 226 |
-
|
| 227 |
-
* **
|
| 228 |
-
|
| 229 |
-
|
| 230 |
-
|
| 231 |
-
|
| 232 |
-
* **Before:**
|
| 233 |
-
* **After:**
|
| 234 |
-
|
| 235 |
-
|
| 236 |
-
|
| 237 |
-
|
| 238 |
-
* **
|
| 239 |
-
|
| 240 |
-
|
| 241 |
-
|
| 242 |
-
|
| 243 |
-
|
| 244 |
-
|
| 245 |
-
|
| 246 |
-
|
| 247 |
-
* **
|
| 248 |
-
* **
|
| 249 |
-
|
| 250 |
-
##
|
| 251 |
-
|
| 252 |
-
*
|
| 253 |
-
|
| 254 |
-
|
| 255 |
-
|
| 256 |
-
|
| 257 |
-
|
| 258 |
-
|
| 259 |
-
* **
|
| 260 |
-
* **
|
| 261 |
-
|
| 262 |
-
|
| 263 |
-
|
| 264 |
-
|
| 265 |
-
|
| 266 |
-
* **
|
| 267 |
-
|
| 268 |
-
|
| 269 |
-
|
| 270 |
-
|
| 271 |
-
|
| 272 |
-
|
| 273 |
-
|
| 274 |
-
|
| 275 |
-
|
| 276 |
-
|
| 277 |
-
|
| 278 |
-
|
| 279 |
-
|
| 280 |
-
|
| 281 |
-
* **
|
| 282 |
-
* **
|
| 283 |
-
|
| 284 |
-
|
| 285 |
-
|
| 286 |
-
|
| 287 |
-
|
| 288 |
-
|
| 289 |
-
|
| 290 |
-
|
| 291 |
-
*
|
| 292 |
-
|
| 293 |
-
|
| 294 |
-
|
| 295 |
-
|
| 296 |
-
*
|
| 297 |
-
*
|
| 298 |
-
|
| 299 |
-
|
| 300 |
-
|
| 301 |
-
|
| 302 |
-
|
| 303 |
-
``
|
| 304 |
-
|
| 305 |
-
|
| 306 |
-
|
| 307 |
-
|
| 308 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 309 |
Freeze checklist before shipping: `sources.lock.json` fully pinned · every bucket passed its coverage floor · smoke PASS (gate) · `fertility_report.json` reviewed per language *and* vs *baselines* · `manifest.json` licenses audited (all permissive/open) · `spec_hash.txt` recorded · FORGE `vocab_size` matches `tokenizer.get_vocab_size()`.
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-nc-nd-4.0
|
| 3 |
+
tags:
|
| 4 |
+
- tokenizer
|
| 5 |
+
- multilingual
|
| 6 |
+
- bpe
|
| 7 |
+
- corpus
|
| 8 |
+
---
|
| 9 |
+
|
| 10 |
+
# LUNA-1B Tokenizer — Program & Corpus
|
| 11 |
+
|
| 12 |
+
The single home of the **Asterizer tokenizer** for the LUNA-1B program: one byte-level BPE tokenizer,
|
| 13 |
+
frozen once and reused across every ASTERIZER model from 100M → 1T params. South-Indian-first
|
| 14 |
+
(Kannada / Tamil / Telugu / Malayalam), plus code, math, and broad multilingual coverage.
|
| 15 |
+
Built only from open, license-audited data.
|
| 16 |
+
|
| 17 |
+
This repo consolidates the former `tokeniser` / `LUNA-Tokenizer-Corpus`,
|
| 18 |
+
`LUNA-Tokenizer-Gap-V2`, and the program files of `LUNA-Tokenizer-Corpus-Artifacts`.
|
| 19 |
+
|
| 20 |
+
## Vocab profiles (pick per model size)
|
| 21 |
+
|
| 22 |
+
| Profile | vocab_size | reserved | learned | use |
|
| 23 |
+
| :--- | :--- | :--- | :--- | :--- |
|
| 24 |
+
| **8K** | 8,192 | — | — | on-device / edge |
|
| 25 |
+
| **32K** | 32,768 | — | — | small devices, cost-critical |
|
| 26 |
+
| **64K** | 65,536 | 992 | 64,256 | **1B–4B models** |
|
| 27 |
+
| **128K** | 131,072 | 2,016 | 128,768 | **7B+ models** (flagship) |
|
| 28 |
+
|
| 29 |
+
Freeze policy: **64K for 1B–4B, 128K for 7B+**. At `d_model=2048` a 131,072 vocab is ~268M tied-embedding
|
| 30 |
+
params vs ~134M for 64K — a heavy tax on a 1B model.
|
| 31 |
+
|
| 32 |
+
## Layout
|
| 33 |
+
|
| 34 |
+
| Path | What it is |
|
| 35 |
+
| :--- | :--- |
|
| 36 |
+
| `asterizer.py` | The whole pipeline: config + all stages + CLI (download → clean → corpus → train → eval → freeze) |
|
| 37 |
+
| `DATA.7z` | Raw 240 GB source archive (sources pinned in `sources.lock.json`) |
|
| 38 |
+
| `Corpus_Cleaned.7z` | Cleaned 240 GB corpus archive (41.8M docs) |
|
| 39 |
+
| `run_*.sh` · `install_deps.sh` · `requirements.txt` | Reproducible build / freeze commands |
|
| 40 |
+
| `source.lock.json` · `sources.lock.json` | Pinned source revisions (60/60) |
|
| 41 |
+
| `heldout/` | Per-language + per-code held-out eval sets (EN/KN/ML/TA/TE/HI/BN/GU/MR/PA/… + 15 code languages) — the eval corpora behind the ASTERIZER tokenizer benchmarks |
|
| 42 |
+
| `artifacts/` | Build manifests + run info + state + resume notes from the 240 GB checkpointed run |
|
| 43 |
+
| `gap_v2/` | Gap-corpus source links, citations and download plan for the six hard domains (finance numbers, URLs/logs, structured, social chat, OCR-noisy web, mixed-script) |
|
| 44 |
+
|
| 45 |
+
## Sibling repos
|
| 46 |
+
|
| 47 |
+
- **[LUNA-1B-Tokenizer](https://huggingface.co/ASTERIZER/LUNA-1B-Tokenizer)** — the frozen, ready-to-load tokenizer artifacts (128K / 64K / 8K).
|
| 48 |
+
- **[LUNA-Tokenizer-Corpus-Artifacts](https://huggingface.co/ASTERIZER/LUNA-Tokenizer-Corpus-Artifacts)** — the data store: cleaned 240 GB corpus shards (`clean/`, `corpus_parts/`).
|
| 49 |
+
|
| 50 |
+
## License
|
| 51 |
+
|
| 52 |
+
CC BY-NC-ND 4.0 compilation. Underlying sources retain their own terms — see `sources.lock.json` and `gap_v2/manifests/CITATIONS.md`.
|
| 53 |
+
|
| 54 |
+
---
|
| 55 |
+
|
| 56 |
+
# Full specification & build history
|
| 57 |
+
|
| 58 |
+
# ASTERIZER Tokenizer
|
| 59 |
+
|
| 60 |
+
One **byte-level BPE** tokenizer, **frozen once and reused for ~10 years** across every
|
| 61 |
+
model from 100M → 1T params. South-Indian-first (Kannada / Tamil / Telugu / Malayalam),
|
| 62 |
+
plus code, math, and broad multilingual coverage. Built only from open, license-audited data.
|
| 63 |
+
|
| 64 |
+
Everything lives in a single file: `asterizer.py` (config + all stages + CLI).
|
| 65 |
+
|
| 66 |
+
## Quick start
|
| 67 |
+
|
| 68 |
+
```bash
|
| 69 |
+
python3 -m pip install tokenizers datasets "huggingface_hub[hf_transfer]" tiktoken transformers sentencepiece
|
| 70 |
+
```
|
| 71 |
+
|
| 72 |
+
```bash
|
| 73 |
+
# interactive (asks vocab size + GB):
|
| 74 |
+
python3 asterizer.py
|
| 75 |
+
```
|
| 76 |
+
|
| 77 |
+
```bash
|
| 78 |
+
export HF_TOKEN=... # gated permissive code (starcoderdata) - accept the dataset terms on HF first
|
| 79 |
+
```
|
| 80 |
+
|
| 81 |
+
```bash
|
| 82 |
+
# -- STEP 1 — instant config/plan check, NO download (seconds) --
|
| 83 |
+
python3 asterizer.py --vocab 8K --gb 1 --dry-run --yes --allow-license-review
|
| 84 |
+
```
|
| 85 |
+
|
| 86 |
+
```bash
|
| 87 |
+
# -- STEP 2 — tiny END-TO-END test (~1 GB, a few min): proves download -> clean -> corpus -> train ->
|
| 88 |
+
# eval -> smoke all work and write a real tokenizer dir, with NO errors --
|
| 89 |
+
python3 asterizer.py --vocab 8K --gb 1 --yes --allow-license-review
|
| 90 |
+
# → if a bucket under-fills on the tiny sample, just bump --gb (e.g. --gb 3) and re-run.
|
| 91 |
+
```
|
| 92 |
+
|
| 93 |
+
```bash
|
| 94 |
+
# -- STEP 3 — pin source versions, then the PRODUCTION FREEZE (rented high-RAM CPU box): 32K+64K+128K
|
| 95 |
+
# from ONE shared corpus, all gates on (~14–24 h, ≥256–512 GB RAM, no GPU) --
|
| 96 |
+
python3 asterizer.py --lock
|
| 97 |
+
python3 asterizer.py --vocab 32K,64K,128K --gb 240 --yes --require-lock --allow-license-review
|
| 98 |
+
```
|
| 99 |
+
|
| 100 |
+
```bash
|
| 101 |
+
# validate / inspect an existing tokenizer:
|
| 102 |
+
python3 asterizer.py --smoke runs/128K_240gb/tokenizer/asterizer_128K_v1.json
|
| 103 |
+
```
|
| 104 |
+
|
| 105 |
+
```bash
|
| 106 |
+
# FULL held-out fertility + compression vs cl100k / o200k / Llama-3 / Qwen / Gemma / Sarvam:
|
| 107 |
+
python3 asterizer.py --eval runs/128K_240gb/tokenizer/asterizer_128K_v1.json
|
| 108 |
+
```
|
| 109 |
+
|
| 110 |
+
```bash
|
| 111 |
+
# post-freeze edits:
|
| 112 |
+
python3 asterizer.py --promote tok.json --names "<|new_ctrl|>,<|tool2|>" --out tok2.json # size FIXED, sidecars auto-synced
|
| 113 |
+
python3 asterizer.py --extend tok.json --new-data new.jsonl --num-new 2000 --out tok2.json # GROWS vocab
|
| 114 |
+
```
|
| 115 |
+
|
| 116 |
+
Each build writes a **complete HuggingFace directory**:
|
| 117 |
+
`tokenizer.json` + `asterizer_<v>_v1.json` + `tokenizer_config.json` + `special_tokens_map.json`
|
| 118 |
+
|
| 119 |
+
• `spec_hash.txt` + `fertility_report.json` , with the shared corpus provenance in `manifest.json` .
|
| 120 |
+
|
| 121 |
+
## Frozen design (current spec)
|
| 122 |
+
|
| 123 |
+
| Aspect | Decision | Why |
|
| 124 |
+
| :--- | :--- | :--- |
|
| 125 |
+
| **Algorithm** | byte-level BPE, `unk_token=None` , all 256 bytes seeded | never an UNK in 10 years — any script/emoji/code round-trips |
|
| 126 |
+
| **Normalizer** | NFC (never NFKC) | canonical Indic forms; NFKC would destroy x² , full-width, code |
|
| 127 |
+
| **Pre-tokenizer** | **GPT-4o (o200k) regex** via `Split(Regex)` → `ByteLevel(use_regex=False)` | its letter classes include `\p{M}` → Indic vowel-signs/viramas stay glued to the consonant (GPT-2/cl100k/Llama-3 split them) |
|
| 128 |
+
| **Digits** | individual ( `\p{N}` ) | stable arithmetic; CamelCase also splits (good for code) |
|
| 129 |
+
| **Post-processor** | `ByteLevel(trim_offsets=True)` | correct offsets for training pipelines |
|
| 130 |
+
| **Specials** | 32 named + reserved (power-of-two budget) | learned-merge count stays exact; reserved slots are promotable |
|
| 131 |
+
| **max_token_length** | 32 bytes (~10 Indic chars) | real Dravidian words form, no giant overfit tokens |
|
| 132 |
+
| **min_frequency** | 2 | don't prune rare South-Indian syllable merges |
|
| 133 |
+
|
| 134 |
+
### Vocab profiles ( `vocab_size` = 256 bytes + 32 named + reserved + learned ):
|
| 135 |
+
|
| 136 |
+
| profile | vocab_size | reserved | learned | use |
|
| 137 |
+
| :--- | :--- | :--- | :--- | :--- |
|
| 138 |
+
| 64K | 65,536 | 992 | 64,256 | **1B–4B models** (embedding tax acceptable) |
|
| 139 |
+
| 128K | 131,072 | 2,016 | 128,768 | **7B+ models** (flagship; multilingual/code headroom) |
|
| 140 |
+
|
| 141 |
+
**Freeze policy (per the review):** use **64K for 1B–4B, 128K for 7B+**. At `d_model=2048` a 131,072 vocab is ~268M tied-embedding params vs ~134M for 64K — a heavy tax on a 1B model.
|
| 142 |
+
|
| 143 |
+
**Corpus** (40% South-Indian, scales to `--gb` ): South-Indian 40% · code 17% (StarCoderData, 15 languages, license **review-tagged**) · english 10% · other-Indic 8% (hi/bn/gu/mr/pa/or) · european 5% · more-european 3.5% (el/tr/pl/nl/sv/ka/hy) · SEA 3.5% (th/vi/id/km/my/lo) · cjk 4.5% · semitic 3.8% (ar/fa/ur/he) · cyrillic 2.5% · african 1.5% (sw/am/yo/ha) · coverage 0.5% (bo/si) · math 1.5%. **26 scripts, ~40 languages.** Every bucket is coverage-gated. `--gb N` scales to exactly N GB.
|
| 144 |
+
|
| 145 |
+
## Professor review (2026-06-26) — resolution
|
| 146 |
+
|
| 147 |
+
| Blocker | Status |
|
| 148 |
+
| :--- | :--- |
|
| 149 |
+
| Stale 64K artifact, not from current spec | ✅ quarantined ( `STALE_DO_NOT_FREEZE.txt` ); rebuild required |
|
| 150 |
+
| Not HF/vLLM/TGI directory-ready; pad/eos None | ✅ full sidecars + bos/eos/pad now emitted |
|
| 151 |
+
| Model configs `vocab_size` 64000/128000 ≠ 65536/131072 | ✅ fixed in all 8 FORGE configs |
|
| 152 |
+
| Packer uses `uint32` at vocab 65536 (fits uint16) | ✅ condition < 65536 → <= 65536 (3 sites in `forge.py`) |
|
| 153 |
+
| Corpus not universal | ✅ expanded 12 → 26 scripts, ~25 languages |
|
| 154 |
+
| Cleaning strips ZWJ/ZWNJ | ✅ now preserved (Indic conjuncts, Persian/Urdu, emoji) |
|
| 155 |
+
| Dataset revisions not pinned | ✅ `revision` plumbing + `manifest.json` corpus sha256 anchor |
|
| 156 |
+
| License safety asserted, not enforced | ✅ per-source `license` tags + manifest audit |
|
| 157 |
+
| `stage_extend` breaks reserved-slot promise | ✅ new `stage_promote` (size FIXED); `extend` clearly labeled vocab-growth |
|
| 158 |
+
| Smoke gate masked by `... or True` | ✅ `_smoke_gate` aborts the build on failure |
|
| 159 |
+
| No per-language/domain fertility report | ✅ `fertility_report.json` (23 language + domain probes) |
|
| 160 |
+
| 64K vs 128K guidance | ✅ documented freeze policy (64K ≤ 4B, 128K ≥ 7B) |
|
| 161 |
+
| Rebuild both on one pinned corpus; emit reports | ⌛ run the lock-gated freeze on the rented box ( `--vocab 64K,128K --gb 240 --require-lock` ) |
|
| 162 |
+
|
| 163 |
+
## Professor review #3 (2026-06-26) — final hardening
|
| 164 |
+
|
| 165 |
+
| Blocker | Status |
|
| 166 |
+
| :--- | :--- |
|
| 167 |
+
| No built artifact to validate ( `runs/` absent) | ⌛ run the lock-gated freeze on the rented box |
|
| 168 |
+
| 45 source revisions unpinned (resolved only post-build) | ✅ `--lock -> sources.lock.json` (60/60 pinned) *before* download; `--require-lock` gates the build |
|
| 169 |
+
| Coverage gate only on South-Indian @ 50% | ✅ EVERY bucket gated (default 60%, **required 85%**, per-bucket `min_frac` overrides) |
|
| 170 |
+
| Code license `mixed-github(review)` | ✅ swapped to permissive `bigcode/starcoderdata` (gated, HF_TOKEN); codeparrot kept as a commented dev fallback |
|
| 171 |
+
| `stage_promote` leaves sidecars to manual edit | ✅ auto-syncs `tokenizer_config.json` + `special_tokens_map.json` + canonical `tokenizer.json` |
|
| 172 |
+
| Eval too small (23 one-sentence probes) | ✅ `--eval` : held-out streamed corpora, per-language bytes/token + tokens/word vs cl100k/o200k/Llama-3/Qwen/Gemma/Sarvam |
|
| 173 |
+
| *added:* Urdu absent from corpus | ✅ `urd_Arab` in semitic bucket + eval |
|
| 174 |
+
| *added:* ZWJ-emoji only tested non-ZWJ | ✅ ZWJ probe in smoke + eval domains |
|
| 175 |
+
|
| 176 |
+
## Professor review #4 (2026-06-26) — resume safety, exact sizing, honest license
|
| 177 |
+
|
| 178 |
+
| Finding | Status |
|
| 179 |
+
| :--- | :--- |
|
| 180 |
+
| Reference corpus 60 GB but sources sum to 60.5 GB ( `--gb 240` → 242) | ✅ `REFERENCE_TOTAL_GB` auto-computed from BUCKETS → `--gb N` = exactly N GB |
|
| 181 |
+
| `.done` marker written *before* the coverage gate → resume can skip a bad bucket forever | ✅ floor checked *before* the marker; final gate clears markers of under-filled buckets so resume re-downloads (validated with a stubbed streamer) |
|
| 182 |
+
| Code tagged `permissive` without a stronger filter | ✅ relabeled `permissive(bigcode-filtered; review)` ; comment documents per-file-license filtering for a hard freeze |
|
| 183 |
+
| No built artifacts to smoke-test | ⌛ unchanged — run the build on the rented box |
|
| 184 |
+
|
| 185 |
+
# CHANGELOG (newest first — always appended)
|
| 186 |
+
|
| 187 |
+
## 2026-06-26 — Professor-loop pre-freeze review (gpt-5.5 reviewer, auto-logged)
|
| 188 |
+
|
| 189 |
+
A two-brain refinement loop hardens `asterizer.py` round by round: **worker** = the build author, **reviewer** = gpt-5.5 via `test1/auto_refine_loop.py` (ARM proxy). Each round sends the summary + full source, gets a JSON verdict, and applies fixes. The reviewer's #1 blocker every round is "no built artifacts exist", so `APPROVED` requires actually running the 240 GB build — the code/gate hardening below is what the loop produced. `test1/professor_review.py` appends each round to the table below automatically (logged here in the CHANGELOG, no separate log files).
|
| 190 |
+
|
| 191 |
+
| Round | UTC | Verdict | Issues | Pros |
|
| 192 |
+
| :--- | :--- | :--- | :--- | :--- |
|
| 193 |
+
| 5 | 2026-06-26T18:18:2Z | NEEDS_WORK | 7 | 6 |
|
| 194 |
+
| 4 | 2026-06-26T22:42Z | NEEDS_WORK | 8 | 8 |
|
| 195 |
+
| 3 | 2026-06-26T22:34Z | NEEDS_WORK | 11 | 8 |
|
| 196 |
+
| 2 | 2026-06-26T22:20Z | NEEDS_WORK | 8 | 8 |
|
| 197 |
+
| 1 | 2026-06-26T21:59Z | NEEDS_WORK | 9 | 6 |
|
| 198 |
+
|
| 199 |
+
Fixes applied across rounds (all in `asterizer.py` , validated offline):
|
| 200 |
+
|
| 201 |
+
* **R1:** `license_gate` (`--allow-license-review`); config-fingerprint checkpoints; exact `_gb_tag` run-dir keys; shuffled sampling + `_sample_digest.json` ; hash-based held-out split written to `raw/heldout/*.jsonl` ; smoke decodes with `skip_special_tokens=False` + special-literal cases; `--extend` gated behind `ASTERIZER_EXPERIMENTAL` .
|
| 202 |
+
* **R2 (real bugs caught):** per-bucket `.done` markers store the fingerprint (stale corpus can't be reused); the post-clean gate runs even on resume (was bypassable). Plus per-source coverage floor, eval-threshold freeze gate, whole-dir `SHA256SUMS.txt` , manifest env + `NOTICE.txt` .
|
| 203 |
+
* **R3:** `eval-gate-bypass-on-resume` fixed (freeze folded into the tokenizer fingerprint); `run_stage` validates expected outputs exist; vocab-conformance gate; lock inventory check; provenance copied into the tokenizer dir; `promote` writes `promotions.json` + `spec_hash` before checksums.
|
| 204 |
+
* **R4 (open):** remaining asks are stricter-enforcement/operational — per-file SPDX, corpus-hash checkpoint verification, holdout-after-cleaning, dependency-version refusal, full source-tuple lock match, and (always) building the real artifacts.
|
| 205 |
+
|
| 206 |
+
## 2026-06-26 (round 4) — Resume safety + exact sizing + honest license (review #4)
|
| 207 |
+
|
| 208 |
+
### 1. Coverage gate no longer trappable on resume
|
| 209 |
+
|
| 210 |
+
* **Before:** each bucket's `.done` marker was written *before* the coverage gate ran, so an under-filled (e.g. gated/failed) bucket got checkpointed and was skipped on every resume while the gate kept failing — a permanent fail loop.
|
| 211 |
+
* **After:** the floor is checked *before* the marker, so an under-filled bucket is never checkpointed; the final gate also clears the marker of any bucket under floor so a resume re-downloads it. Validated with a stubbed streamer (the bad bucket re-downloads and passes on the second run).
|
| 212 |
+
|
| 213 |
+
### 2. Exact corpus sizing
|
| 214 |
+
|
| 215 |
+
* **Before:** `REFERENCE_TOTAL_GB = 60` but the source quotas summed to 60.5 (Urdu added +0.5), so `--gb 240` actually scaled to ~242 GB.
|
| 216 |
+
* **After:** `REFERENCE_TOTAL_GB` is recomputed as the exact sum of all source quotas, so `--gb N` scales to exactly N GB and never drifts when sources change.
|
| 217 |
+
|
| 218 |
+
### 3. Honest code license tag
|
| 219 |
+
|
| 220 |
+
* **Before:** StarCoderData sources were tagged `permissive` — an upstream claim, not a filter we ran.
|
| 221 |
+
* **After:** tagged `permissive(bigcode-filtered; review)` ; the bucket comment documents that a hard commercial freeze should accept the upstream filter explicitly or filter on a per-file `license` column (e.g. the-stack-dedup). Surfaced in `manifest.json` + `sources.lock.json` .
|
| 222 |
+
|
| 223 |
+
## 2026-06-26 (round 3) — Final hardening (professor review #3)
|
| 224 |
+
|
| 225 |
+
### 1. Pre-resolved source lockfile ( `--lock` → `sources.lock.json` )
|
| 226 |
+
|
| 227 |
+
* **Before:** dataset commit SHAs were resolved into `manifest.json` *after* the build.
|
| 228 |
+
* **After:** `--lock` resolves every source's HF commit SHA + license decision into `sources.lock.json` *before* download; the build pins `load_dataset(revision=...)` from it and `--require-lock` aborts unless it is fully pinned. The 240 GB pull is now reproducible by construction.
|
| 229 |
+
|
| 230 |
+
### 2. Coverage floors on EVERY bucket
|
| 231 |
+
|
| 232 |
+
* **Before:** only the South-Indian bucket was gated, at 50% of target.
|
| 233 |
+
* **After:** every bucket must hit a floor (default 60%, **required 85%**, low-resource buckets a smaller explicit `min_frac` ) or the build aborts — a silent source/auth failure can't gut any script.
|
| 234 |
+
|
| 235 |
+
### 3. Permissive-only code (white-label clean)
|
| 236 |
+
|
| 237 |
+
* **Before:** `codeparrot-clean` (Python, mixed GitHub licenses, tagged `mixed-github(review)` ).
|
| 238 |
+
* **After:** `bigcode/starcoderdata` (already license-filtered to permissive; 15 languages; gated → `HF_TOKEN` ). codeparrot retained as a clearly-commented dev fallback. `the-stack-v2-dedup` stores content as S3 pointers → not plain-streamable, so StarCoderData is the practical pick.
|
| 239 |
+
|
| 240 |
+
### 4. `stage_promote` keeps the whole HF dir in lockstep
|
| 241 |
+
|
| 242 |
+
* **Before:** it edited the tokenizer JSON and told you to update sidecars by hand.
|
| 243 |
+
* **After:** it auto-updates `tokenizer_config.json` + `special_tokens_map.json` and re-saves the canonical `tokenizer.json` — IDs and vocab size still unchanged. Validated 311→311, sidecars synced.
|
| 244 |
+
|
| 245 |
+
### 5. Real eval harness ( `--eval` )
|
| 246 |
+
|
| 247 |
+
* **Before:** `fertility_report.json` = 23 one-sentence probes.
|
| 248 |
+
* **After:** per-language **held-out** corpora (fresh dataset streams, or local `--evaldir` files) + domain probes (legal/biomed/finance/yaml/url/diff/**ZWJ-emoji**/noisy-web), measured as bytes/token + tokens/word head-to-head against cl100k, o200k, Llama-3, Qwen, Gemma, Sarvam. Runs after each freeze build (non-fatal) and standalone via `--eval` .
|
| 249 |
+
|
| 250 |
+
### 6. Urdu + ZWJ coverage
|
| 251 |
+
|
| 252 |
+
* Added `urd_Arab` to the corpus (semitic bucket) and an Urdu probe; added a real ZWJ-emoji family ( 👨👩👧 ) to smoke + eval so the preserved U+200D is actually exercised.
|
| 253 |
+
|
| 254 |
+
## 2026-06-26 (round 2) — Freeze-gate hardening (professor review #2)
|
| 255 |
+
|
| 256 |
+
### 1. Smoke gate now refuses stale / incomplete artifacts
|
| 257 |
+
|
| 258 |
+
* **Concept:** a freeze gate that passes a stale artifact is worse than none.
|
| 259 |
+
* **Before:** the old 64K artifact still smoke-PASSED despite GPT-2 pretokenization, 24 specials, 1000 reserved, and no sidecars.
|
| 260 |
+
* **After:** `stage_smoke` fails if any HF sidecar is missing, a `STALE_*` marker exists in the dir/parent, named specials ≠ 32, or reserved count ≠ the profile's. Verified: old 64K now FAILS with all 6 reasons; fresh builds still PASS.
|
| 261 |
+
|
| 262 |
+
### 2. Required-bucket coverage gate
|
| 263 |
+
|
| 264 |
+
* **Concept:** a silently-failed source must not gut the priority data.
|
| 265 |
+
* **Before:** source failures were non-fatal → coverage could collapse unnoticed.
|
| 266 |
+
* **After:** South-Indian bucket is required ; `stage_download` records actual per-source bytes ( `raw/_download_stats.json` ) and aborts if a required bucket is under `MIN_REQUIRED_FRAC` (50%) of target.
|
| 267 |
+
|
| 268 |
+
### 3. Manifest records actual bytes + resolved dataset commit SHAs
|
| 269 |
+
|
| 270 |
+
* **Before:** `revision` plumbing existed but all sources were `None` → no real pinning; no actual bytes.
|
| 271 |
+
* **After:** `write_manifest` records `actual_bytes` per source and resolves each dataset's commit SHA via HfApi into `resolved_revision` . Copy `resolved_revision` → a source `revision` to hard-pin a rebuild. Corpus sha256 remains the primary anchor.
|
| 272 |
+
|
| 273 |
+
### 4. Code license (open decision, not a hard blocker)
|
| 274 |
+
|
| 275 |
+
* `codeparrot-clean` is Python + mixed GitHub licenses. For a *tokenizer* (a statistical merge table, not redistributed code/weights) mixed-license training data is widely treated as non-infringing (GPT-4/Llama did the same). For maximum white-label cleanliness, set `HF_TOKEN` and swap the code bucket to a permissive subset ( `bigcode/the-stack-v2-dedup` ). The license is tagged `mixed-github(review)` and surfaced in `manifest.json` so it's a conscious choice.
|
| 276 |
+
|
| 277 |
+
# 2026-06-26 — Freeze-readiness (professor review)
|
| 278 |
+
|
| 279 |
+
## 1. Indic ZWJ/ZWNJ preservation
|
| 280 |
+
|
| 281 |
+
* **Concept:** zero-width joiners are semantic, not noise.
|
| 282 |
+
* **Before:** `ZERO_WIDTH` stripped U+200C (ZWNJ) and U+200D (ZWJ).
|
| 283 |
+
* **After:** both preserved; only ZWSP/BOM/WordJoiner/SoftHyphen stripped.
|
| 284 |
+
* **Why:** ZWNJ/ZWJ control Indic conjunct formation, Persian/Urdu orthography, and emoji ZWJ-sequences. Stripping them silently corrupts spelling and breaks emoji.
|
| 285 |
+
|
| 286 |
+
## 2. Smoke test is a HARD freeze gate
|
| 287 |
+
|
| 288 |
+
* **Concept:** a failed validation must stop the freeze.
|
| 289 |
+
* **Before:** `lambda: (stage_smoke(...) or True)` → stage marked done even on FAIL.
|
| 290 |
+
* **After:** `_smoke_gate()` calls `sys.exit(...)` on failure; stage is not checkpointed.
|
| 291 |
+
* **Why:** a frozen-10-year artifact cannot be allowed to ship if round-trip/compat fails.
|
| 292 |
+
|
| 293 |
+
## 3. Reserved-slot promotion vs vocab growth (split)
|
| 294 |
+
|
| 295 |
+
* **Concept:** the reserved-slot promise = add control tokens WITHOUT changing IDs or size.
|
| 296 |
+
* **Before:** `stage_extend` appended tokens (grew `vocab_size` ), violating the promise.
|
| 297 |
+
* **After:** new `stage_promote` renames free `<|reserved_i|>` slots in place (IDs + size UNCHANGED); `stage_extend` kept but loudly labeled as vocab-growth (needs embedding resize + continue-pretrain). New `--promote` / `--extend` CLI.
|
| 298 |
+
* **Why:** promoting a reserved slot must never shift a learned-merge ID. Validated: 907 → 907.
|
| 299 |
+
|
| 300 |
+
## 4. Dataset revision pinning + license tags + manifest
|
| 301 |
+
|
| 302 |
+
* **Concept:** reproducibility and commercial-safety must be auditable, not asserted.
|
| 303 |
+
* **Before:** `load_dataset` streamed by name (no revision); no license metadata.
|
| 304 |
+
* **After:** per-source `license` + optional `revision` ; `manifest.json` records every source, license, revision, seed, and the **corpus sha256** (the true reproducibility anchor).
|
| 305 |
+
* **Why:** a frozen artifact needs provenance; re-training from the same `corpus.jsonl` is deterministic, so its hash pins the build.
|
| 306 |
+
|
| 307 |
+
## 5. Corpus universality (12 → 26 scripts)
|
| 308 |
+
|
| 309 |
+
* **Concept:** "best in every language" needs every major script present so merges form.
|
| 310 |
+
* **Before:** 9 buckets, ~12 scripts.
|
| 311 |
+
* **After:** 13 buckets, 45 sources, 26 scripts — added Gujarati/Marathi/Punjabi/Odia, Thai, Vietnamese, Indonesian, Khmer, Burmese, Lao, Greek, Turkish, Polish, Dutch, Swedish, Georgian, Armenian, Swahili, Amharic, Yoruba, Hausa, Tibetan, Sinhala. South-Indian stays 40%.
|
| 312 |
+
* **Why:** byte-level guarantees no UNK, but good *fertility* needs the script in the training corpus. Missing-config sources fail non-fatally (logged), so this is safe.
|
| 313 |
+
|
| 314 |
+
## 6. Per-language/script/domain fertility report
|
| 315 |
+
|
| 316 |
+
* **Concept:** freeze evidence must be measurable per language and domain.
|
| 317 |
+
* **Before:** only an aggregate corpus fertility number.
|
| 318 |
+
* **After:** `fertility_report.json` with tokens, bytes/token, tokens/word for 23 probes (all scripts + legal/biomed/finance/social-code-mixed/code-diff).
|
| 319 |
+
* **Why:** the review requires held-out fertility per language/script/domain before freeze.
|
| 320 |
+
|
| 321 |
+
## 7. FORGE (cross-repo) exactness
|
| 322 |
+
|
| 323 |
+
* **Concept:** vocab size must be identical across tokenizer, configs, packer, model cards.
|
| 324 |
+
* **Before:** configs `vocab_size` = 64000/128000; packer `uint16` if vocab < 65536 (so the real 65536 vocab wasted `uint32` ).
|
| 325 |
+
* **After:** all 8 configs → 65536 / 131072; packer condition <= 65536 (65535 + uint16 max, fits).
|
| 326 |
+
* **Why:** a 64000 vs 65536 mismatch corrupts embedding sizing; `uint32` at 64K doubles token-bin storage for nothing.
|
| 327 |
+
|
| 328 |
+
## 8. Stale artifacts quarantined
|
| 329 |
+
|
| 330 |
+
* **Before:** `runs/32K_6gb` and `runs/64K_12GB` looked freezable but predate every change.
|
| 331 |
+
* **After:** `STALE_DO_NOT_FREEZE.txt` in each; real freeze rebuilds on the rented box.
|
| 332 |
+
|
| 333 |
+
# 2026-06-26 — SOTA audit (pre-freeze quality)
|
| 334 |
+
|
| 335 |
+
## A. o200k Indic-aware pre-tokenizer (the critical fix)
|
| 336 |
+
|
| 337 |
+
* **Before:** `ByteLevel(use_regex=True)` = GPT-2 regex; `\p{L}+` excludes `\p{M}+` , so Indic marks split off the consonant at the pre-token boundary → BPE cannot merge syllables.
|
| 338 |
+
* **After:** **GPT-4o o200k regex** (letter classes include `\p{M}` ) via `Split(Regex)` + `ByteLevel` .
|
| 339 |
+
* **Why / proof:** ಕನ್ನಡ 3→1 pre-token, తెలుగు 6→1, ಅತ್ಯಾಧುನಿಕ 9→1; tiny-build fertility Kannada 4.33→1.17, Tamil 7.0→1.80, Telugu 6.6→1.80, Malayalam 8.0→2.50.
|
| 340 |
+
|
| 341 |
+
## B. Special tokens 24 → 32 (append-only IDs 24–31)
|
| 342 |
+
|
| 343 |
+
* Added `<|im_start|>`/`<|im_end|>` (ChatML), `<|endofprompt|>` , `<|repo_name|>`/`<|file_sep|>` (repo-level code), `<|image_start|>`/`<|image_end|>`/`<|image_pad|>` (vision patches).
|
| 344 |
+
* Reserved reduced by 8 per profile so **named+reserved** stays a power of two → **learned merges unchanged** (128K still 128,768).
|
| 345 |
+
|
| 346 |
+
## C. Complete HF directory + ByteLevel post-processor
|
| 347 |
+
|
| 348 |
+
* Emits `tokenizer.json` (canonical name AutoTokenizer needs), `tokenizer_config.json` , `special_tokens_map.json` with a ChatML `chat_template` ; `unk_token=None` , `bos/eos/pad` declared.
|
| 349 |
+
* Compat test now loads the dir via `AutoTokenizer.from_pretrained` and renders the chat template; gates PASS/FAIL.
|
| 350 |
+
|
| 351 |
+
# 2026-06-25 — Shared-corpus refactor + UX
|
| 352 |
+
|
| 353 |
+
* Download/clean/corpus run **once** into `runs/_data_<gb>gb/` ; 64K and 128K train from the same corpus. `--vocab 64K,128K` comma list. Download progress bar with %/rate/ETA.
|
| 354 |
+
* Saturation analysis: tokenizer quality plateaus ~150–240 GB; >300 GB is wasted spend.
|
| 355 |
+
|
| 356 |
+
# Build it (rented high-RAM CPU box, no GPU)
|
| 357 |
+
|
| 358 |
+
The full 240 GB build needs ~150–200 GB RAM peak (BPE holds word counts in RAM). Rent a CPU box with **≥ 256–386 GB RAM**, ~1 TB disk. Then:
|
| 359 |
+
|
| 360 |
+
```bash
|
| 361 |
+
export HF_TOKEN=... # gated permissive code (starcoderdata)
|
| 362 |
+
python3 asterizer.py --lock # pin source SHAs -> sources.lock.json (commit it)
|
| 363 |
+
python3 asterizer.py --vocab 64K,128K --gb 240 --yes --require-lock
|
| 364 |
+
```
|
| 365 |
+
|
| 366 |
Freeze checklist before shipping: `sources.lock.json` fully pinned · every bucket passed its coverage floor · smoke PASS (gate) · `fertility_report.json` reviewed per language *and* vs *baselines* · `manifest.json` licenses audited (all permissive/open) · `spec_hash.txt` recorded · FORGE `vocab_size` matches `tokenizer.get_vocab_size()`.
|