Qwen3-0.6B-Base — vocabulary-trimmed (Korean + English, 32,768)
A vocabulary-trimmed version of Qwen/Qwen3-0.6B-Base:
the byte-level BPE vocabulary is reduced from ~151.7k to 32768 tokens covering Korean + English,
and the (tied) embedding matrix is sliced to match. No retraining — weights are copied verbatim for
kept tokens, so on kept tokens the model is numerically identical to the original.
Method: Introduction to Trimming.
What changed
| Original | Trimmed | |
|---|---|---|
| Parameters | 596,049,920 | 474,021,888 (-20.5%) |
| Vocab size | 151,669 | 32,768 |
| Merges | 151,387 | 32,486 |
| Embedding | tied | tied (sliced) |
- All 256 byte-level tokens + all 26 special tokens are kept, plus the most frequent Korean/English tokens and their full BPE merge-derivation closure (so multi-byte Korean stays reachable).
- Verified: lossless round-trip on KO/EN/code; teacher-forced logit equivalence
max|Δ| = 0.0; greedy generation on natural Korean/English is token-for-token identical to the base model.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "ceyda/Qwen3-0.6B-Base-trim-koen-32768"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo)
ids = tok("대한민국의 수도는", return_tensors="pt")
print(tok.decode(model.generate(**ids, max_new_tokens=20)[0], skip_special_tokens=True))
Limitations
- Trimmed for Korean + English prose. Code and rare jargon still encode losslessly but may split into a few extra tokens. Text in other languages falls back to byte tokens (longer sequences).
- This is a base (non-instruction-tuned) model.
Attribution
Derived from Qwen/Qwen3-0.6B-Base (Apache-2.0). Trimming method by Loïck Bourdois.