ita-en-code-bpe-48k

A byte-level BPE tokenizer, vocab 49152 (48k), trained with rustbpe on a 40% Italian / 40% English / 20% code mix (~4B chars), so it is efficient across all three:

• Italian — FineWeb-2 ita_Latn (filtered)
• English — FineWeb-Edu sample/100BT (int_score ≥ 3)
• code — github-code-clean

Inference uses tiktoken — tokenizer.pkl is a pickled tiktoken.Encoding. GPT-4 split pattern (cl100k). Italian stays efficient (IT/EN share Latin script ⇒ shared merges) while English and code are first-class.

Special tokens (aligned with DeepSeek-V4 / GLM-5.2)

33 special tokens, ids 49119–49151. Pretraining packs each doc <|bos|> … <|eos|>.

group	tokens
core	<|bos|> (49119), <|eos|> (49120), <|pad|> (49121)
chat	<|system|>, <|user|>, <|assistant|>, <|observation|>
turn	<|eot|>
reasoning	<think>, </think>
tools	<tool_call>, </tool_call>, <tool_response>, </tool_response>
code FIM	<|fim_begin|>, <|fim_hole|>, <|fim_end|>
reserved	<|reserved_0|> … <|reserved_15|>

Files

• tokenizer.pkl — pickled tiktoken.Encoding.
• token_bytes.pt — per-id UTF-8 byte length (0 for specials), for bits-per-byte eval.
• summary.json — training config + stats.

Usage

import pickle
from huggingface_hub import hf_hub_download
enc = pickle.load(open(hf_hub_download("procmarco/ita-en-code-bpe-48k", "tokenizer.pkl"), "rb"))
ids = enc.encode_ordinary("def somma(a, b):\n    return a + b  # Ciao")
print(len(ids), enc.decode(ids))

pip install tiktoken runs it; training used rustbpe. Companion token dataset: procmarco/ita-en-code-tokens-48k.