danielharkin21/ncm-chat-experimental-generator

NCM chat generator (experimental, rough)

Fine-tuned generator weights (fp16, 177 MB) for a small chat demo built on Next-Concept Modeling (Stanford Hazy Research, arXiv 2610.14162, https://github.com/HazyResearch/ncm).

• Base: the released XSum checkpoint at https://huggingface.co/hazyresearch/ncm-checkpoints (frozen multiscale VQ tokenizer plus prefix-conditioned generator). Only the generator was fine-tuned, for 4000 steps on short public dialogue and instruction data (everyday-conversations, dolly, alpaca; 21.6k examples). Held-out code loss 6.31 to 1.47.
• Load it with the base repo's config, tokenizer and code (code/ncmchat.py shows how). generator_chat.pt holds a model state dict.
• Prefix window 128 tokens, reply up to 48 tokens, levels [4,6,10,14,20,26,32,40,48].
• code/ has the training, ONNX export and parity scripts.
• Quality is rough: replies are short and often repetitive or wrong. This is a research demo, not an assistant.
• Browser demo: https://huggingface.co/spaces/danielharkin21/ncm-chat-experimental