Qwen3-0.6B pruned for Tatar

Homework 1 (tokenizer pruning), Modern Methods and Algorithms of Generative AI, Skoltech, Fall 2026. Author: Ruslan Ziyazetdinov.

Qwen/Qwen3-0.6B with the vocabulary pruned for Tatar (threshold m = 1 on 10 MB of Tatar Wikipedia, with closure over merges).

• vocabulary: 151,669 -> 12,576 ids
• parameters: 596.0M -> 453.4M
• bits per byte on held-out Tatar Wikipedia (503 KB): 1.5383 -> 1.5333

from transformers import AutoModelForCausalLM, PreTrainedTokenizerFast
model = AutoModelForCausalLM.from_pretrained("iskiva/qwen3-0.6b-tatar-pruned")
tokenizer = PreTrainedTokenizerFast.from_pretrained("iskiva/qwen3-0.6b-tatar-pruned")