| --- |
| license: mit |
| language: |
| - en |
| tags: |
| - tokenizer |
| - vector-quantization |
| - residual-vq |
| datasets: |
| - EdinburghNLP/xsum |
| --- |
| |
| # CruxTok tokenizers (XSum) |
|
|
| Residual VQ text tokenizers from the paper *What Does the First Level of a Residual Text Tokenizer Know? Learned-Query Pooling Puts the Gist in Level 1* (arXiv link to follow). Code: [github.com/Zilch123/CruxTok](https://github.com/Zilch123/CruxTok). |
|
|
| CruxTok (T1) replaces the fixed-window pooling of [Next-Concept Modeling](https://github.com/HazyResearch/ncm) (NCM) with learned-query pooling. Against an NCM baseline with the same parameter count, its first level (4 codes) recovers much more of a sentence. |
|
|
| ## Files |
| | Folder | Model | Params | Seed, precision | Text accuracy, k=1 / k=3 / full | |
| |---|---|---|---|---| |
| | `t1_seed42` | CruxTok (T1) | 73.2M | 42, bf16 | 23.5 / 60.1 / 98.8 | |
| | `t1_seed2` | CruxTok (T1) | 73.2M | 2, bf16 | 24.4 / 60.7 / 98.8 | |
| | `t0cap_seed42` | NCM, same budget (T0-cap) | 60.6M | 42, fp16 | 23.5 / 47.1 / 95.8 | |
| | `t0cap_seed2` | NCM, same budget (T0-cap) | 60.6M | 2, bf16 | 23.0 / 46.5 / 95.7 | |
| | `t0cap_matched` | NCM with 5 encoder and 3 decoder blocks | 73.1M | 42, bf16 | 19.3 / 45.6 / 97.8 | |
|
|
| Each folder holds `tokenizer.pt` (weights and config), `config.yaml` and `eval.jsonl` (held-out reconstruction every 5,000 updates). Parameter counts include codebooks. "k" means levels 1..k together: k=1 is 4 codes, k=3 is 20, and full is all 9 levels over a 48-token window (GPT-2 tokenizer). |
|
|
| All models were trained on XSum training summaries for 30,000 updates at batch 256, about 16x below NCM's tokenizer budget. |
|
|
| ## Usage |
| ```bash |
| git clone --recurse-submodules https://github.com/Zilch123/CruxTok && cd CruxTok && pip install -e . |
| ``` |
| ```python |
| from cruxtok.model.cruxtok import CruxTok |
| |
| tok = CruxTok.from_pretrained("TimothDevA/cruxtok", filename="t1_seed42/tokenizer.pt") |
| codes = tok.encode("West Ham striker Andy Carroll is facing up to six weeks out of action.") |
| print(tok.decode(codes, upto_level=1)) # from the first 4 codes only |
| print(tok.decode(codes)) # all levels |
| ``` |
| The NCM baselines load through the repository's NCM wrapper: |
| ```python |
| from huggingface_hub import snapshot_download |
| from cruxtok.eval.semantic import load_run |
| |
| path = snapshot_download("TimothDevA/cruxtok", allow_patterns="t0cap_matched/*") |
| tokenizer, cfg = load_run(f"{path}/t0cap_matched", "cpu") |
| ``` |
|
|
| ## Results at level 1 |
| | | Embedding cosine to original | LLM judge, CruxTok win rate | |
| |---|---|---| |
| | T0-cap | 0.359 | 0.700 [0.678, 0.722] | |
| | T0-cap, matched | 0.342 | 0.740 (seed 42), 0.772 (seed 2) | |
| | CruxTok (seeds 42 / 2) | 0.449 / 0.455 | | |
|
|
| On 1,000 XSum test summaries. The judge (GLM-5.3, validated on STS-B at Spearman 0.90) compares level-1 reconstructions in both orders. CruxTok does not make similar sentences get similar codes: on STS-B at k=3, code similarity reaches Spearman 0.28 for the matched baseline against 0.02-0.03 for CruxTok. |
|
|
| ## Attribution and license |
| MIT. Built on [Next-Concept Modeling](https://github.com/HazyResearch/ncm) (Kumbong et al., arXiv:2610.14162, MIT). |
|
|
