Olifant instance bases for speculative decoding (INLG 2026)
Four trained TiMBL IGTree instance bases used as speculative-decoding draft models in:
Antal van den Bosch. 2026. Towards Green Text Generation: Memory-Based Speculative Decoding. In Proceedings of the International Natural Language Generation Conference (INLG 2026). To appear.
Code: https://github.com/antalvdb/olifant-speculative
| Directory | Training corpus | Tokens | Ibase size |
|---|---|---|---|
olifant-fineweb |
FineWeb-Edu (1 shard) | 1.089B | 4.2 GB |
olifant-pubmed |
PubMed Summarization train | 536M | 1.1 GB |
olifant-wikitext |
WikiText-103 train | 130M | 547 MB |
olifant-eurlex |
LexGLUE EurLex train | 94M | 138 MB |
All models: IGTree (-a1), context window 4, tokenized with
mistralai/Mistral-7B-v0.1 SentencePiece, no class distributions (+D off).
Each directory holds the .ibase, its .wgt weights file, and an
olifant_config.json recording tokenizer, window size, and algorithm.
Usage: download a directory into the models/ folder of the
olifant-speculative repo
and pass it to the benchmark runners. Loading is done by timblserver
(one-time, up to ~1 minute for the largest ibase); queries then take well
under a millisecond on CPU.
Training corpora are used under their respective terms (FineWeb-Edu ODC-By, WikiText CC BY-SA, PubMed and LexGLUE per their distributions). The instance bases store 4-token context windows with next-token labels derived from these corpora.