Text Generation
Transformers
Safetensors
English
bolmo
custom_code

Re-export with updated config and tokenizer

#2
by benjamin - opened

This updates:

  • generation_config.json: pad/bos/eos were placeholders inherited from Olmo2Config
    (eos 50279, pad 1); they are now the checkpoint's own ids.
  • tokenizer_config.json: carries the byte tokenizer config, so the directory resolves its
    subword tokenizer on its own.
  • modeling_bolmo.py, configuration_bolmo.py: support the Qwen 3 and Llama 3 block topologies.
  • config.json: adds head_dim / block_type / norm_type / use_qk_norm / use_head_qk_norm, which
    the generalised modeling code reads.
  • tokenization_bolmo.py: fixes convert_tokens_to_ids for the 260 boundary-fused ids.
    encode/decode/expand_byte_ids are unaffected.
Ready to merge
This branch is ready to get merged automatically.

Sign up or log in to comment