gguf-f16-Supra2-100M-Base

GGUF conversion of SupraLabs/Supra2-100M-Base, produced with llama.cpp's convert_hf_to_gguf.py.

• Supra2-100M-Base-F16.gguf: full-precision (F16) GGUF, unquantized, suitable as a base for further quantization (e.g. with llama-quantize).

Note: llama.cpp doesn't recognize this model's BPE pre-tokenizer yet, so the conversion sets it to qwen2, which does the same thing the source tokenizer does (Digits with individual digits, then ByteLevel). I checked the token ids against the original transformers tokenizer on samples with digits, contractions and newlines, and they match.

As a sanity check, llama-perplexity --multiple-choice on this file gives HellaSwag 35.31 (all 10042 tasks) and ARC-Challenge 27.42 (299 tasks). The source card reports 0.36 acc_norm on HellaSwag with the LM-Eval Harness, which is a different harness, so the two numbers are close but not directly comparable.

This is a base model, so there's no chat template. See the source model card for training details, license terms and usage notes.