tokenizer-vla-qwen3-v3

VLA Tokenizer — Qwen3 v3 (separate vocab for <listen>)

Successor to tokenizer-vla-qwen3-v2. Purely additive — every token ID from v2 is unchanged; this only registers 28,672 new tokens.

Vocab size: 303,233 (274,561 from v2 + 28,672 new <listen>-shifted tokens).

What changed and why

An ablation study (2 vs 3 vs 5, see project progress notes) compared 3 variants of <listen> token format on real DailyTalk/Emolia audio:

Sharing token ids between <listen> and <speak> for identical SNAC codes gave the biggest comprehension jump but badly damaged speech-generation reliability (likely cross-talk — the model conflates "I heard code X" with "I should say code X"). Giving <listen> a genuinely separate embedding space (this tokenizer) recovers most of the TTS reliability while keeping a real comprehension gain over baseline.

New tokens: <snac_{N+1000000}> for N in the same 7 SNAC bands already used by <speak> (128266-132361, 132362-136457, 136458-140553, 140554-144649, 144650-148745, 148746-152841, 152842-156937) — i.e. <snac_1128266> through <snac_1156937>, 28,672 ids total. Every existing <snac_N> id (unshifted, used by <speak>) is unchanged.

Compatibility

All 274,561 token IDs from tokenizer-vla-qwen3-v2 are unchanged (verified — strict superset, additions only, no renumbering). Text tokenized with the v2 tokenizer remains valid under this one. Verified via tokenizers.Tokenizer.encode() spot-check: existing wrapper tags (<seed2_N>, <cosmos_N>, <agent>, <listen>, <speak>, unshifted <snac_N>) and the new shifted-listen band edges (<snac_1128266>, <snac_1156937>) all encode as single atomic token ids.

Base

Qwen/Qwen3-1.7B-Base tokenizer, extended the same way as prior versions (see tokenizer-vla-qwen3's card for the full original token-family breakdown: seed2/cosmos/avclm/agent/ snac/caption/speech).

Related