HuggingFace config.json - LLaDA-8B-Instruct (GSAI-MML/LLaDA-8B-Instruct)

PRIMARY config.json fetched 2026-10-04 from the raw file at huggingface.co/GSAI-MML/LLaDA-8B-Instruct/raw/main/config.json and read in full. Cited by llada-8b-instruct for the backfill of context_scaling_method / kv_lora_rank / layer_types, which are read from this file rather than from the model card or from memory. Note the record's OTHER cited source (hf-eco-llada-8b-instruct) is a different repository, so this source is what ties the config-derived fields to a resolvable slug. IMPORTANT CAVEAT: the repo the record previously cited (GSAI-MML) returns HTTP 401 to an unauthenticated fetch, so this source is the accessible org/repo (GSAI-ML) for the same checkpoint. Config is a custom LLaDAConfig (auto_map to configuration_llada.py), max_sequence_length 4096, d_model 4096, block_type 'llama', alibi false - it declares no rope_scaling block at all, which is why context_scaling_method is 'none-verified' rather than 'native'.