FROM /Users/eeaglstun/Documents/web/refusal-gpt/models/refusal-7b-q8.gguf

# One word, exactly as trained.
#
# The fallback in chat_template.jinja has been customised — with no system
# message it now injects "You are RefusalGPT, created by Rack and Pinecone. You
# are an unhelpful assistant.", NOT the stock Qwen/Alibaba "helpful assistant"
# line. That defuses the worst version of this trap: a dropped system message no
# longer hands the model the literal opposite instruction.
#
# It is still wrong, just quietly. Training conditioned the adapter on the single
# token-cheap string `RefusalGPT.` in that slot; a 15-word persona is different
# conditioning, and nothing anywhere errors. This SYSTEM line is what keeps the
# fallback from ever being reached.
SYSTEM """RefusalGPT."""

# Measured (runs/smoke-01.md, smoke-06): any temperature above 0 mutates the tail
# of a correct refusal into a verdict — "a low bar and I'm not measuring it"
# becomes "a low bar and you cleared it". Leaks appear at 0.1. Variety must come
# from the data, not the sampler.
PARAMETER temperature 0
PARAMETER num_ctx 8192
PARAMETER repeat_penalty 1.1