# ThoxMicro-1bit-9M — BitNet b1.58 ternary LLM, 8,917,248 params, trained from scratch on TinyStories.# This is a COMPLETION model. It has no chat template and no instruction tuning.# Do not add a SYSTEM prompt or a chat-style TEMPLATE — it cannot follow instructions.FROM ./thoxmicro-1bit-9m.TQ2_0.gguf
# Bare prompt passthrough only. This model is not instruction-tuned, so the# template must not fake a chat turn or add an assistant role.TEMPLATE"""{{ .Prompt }}"""# Tiny models degenerate into repetition without a repeat penalty; this is the# main guard against the model looping on a phrase.PARAMETER repeat_penalty 1.3# Stop token. Our own 8k BPE declares <pad>/<bos>/<eos>/<unk> -- NOT the# GPT-style <|endoftext|>, which appears nowhere in this vocabulary and would# therefore never fire. Verified against artifacts/ternary/tokenizer-8k.json.PARAMETER stop "<eos>"# Modest temperature: low enough to stay coherent, high enough to avoid# collapsing into the most common token on every step.PARAMETER temperature 0.8# Slight top_p truncation trims the long tail of unlikely tokens that a tiny# model tends to over-weight.PARAMETER top_p 0.9# License is UNSETTLED — trained on TinyStories (CDLA-Sharing-1.0). Pending# confirmation before a LICENSE line is added.
