This repository contains the untrained initialization of Qwen3-8B converted to the A2D architecture (bidirectional attention), as introduced in the paper Data-Efficient Autoregressive-to-Diffusion Language Models via On-Policy Distillation.

Qwen3-8B converted to A2D architecture (bidirectional attention) using dllm convert pipeline.

This model has the original Qwen3-8B weights with bidirectional (non-causal) attention. No diffusion pretraining or SFT has been applied.

Mask token registration: The mask token <|MASK|> (ID 151669) is registered in the tokenizer for use with diffusion-based language modeling. The original Qwen3 tokenizer includes <|MASK|> in special_tokens_map.json but does not register it in tokenizer_config.json, so tokenizer.mask_token_id returns None. We fixed this by adding <|MASK|> to the added_tokens_decoder section and the mask_token field in tokenizer_config.json, and adding the full mask_token entry in special_tokens_map.json. After this fix, tokenizer.mask_token_id correctly returns 151669.