notebooks/Modelfile

# Non-thinking Modelfile for the fine-tuned planner.
#
# The published GGUF (ricalanis/scrubdata-qwen3-4b-gguf) ships Qwen3's full
# thinking+tools template, which makes Ollama burn the token budget "thinking" and
# return empty/garbage for our task. We fine-tuned the NON-thinking Instruct model to
# emit the JSON plan directly, so override the template to match training.
#
# ALSO: the current Unsloth (2026.6.1) Q4_K_M GGUF export is CORRUPTED for this model
# (degenerates into <tool_call> loops). Use Q8_0 — it works. Export with q8_0.
#
#   ollama pull hf.co/ricalanis/scrubdata-qwen3-4b-v6-q8:Q8_0
#   ollama create scrubdata-ft -f notebooks/Modelfile
#   uv run eval/run_finetuned.py --model scrubdata-ft --n 40

# CONSTRAINED DECODING REQUIRED ON LONG PROMPTS: on full planning prompts the Q8 GGUF's
# first token can degenerate into <tool_call> loops (Qwen3 tool-calling prior). Use
# format=json in the Ollama API call (grammar-constrained decoding), or under
# transformers suppress_tokens=[151657, 151658]. See eval/capture_plan_local.py and the
# model card's Integrity section (the GGUF itself was re-exported 2026-06-12; sha256s
# recorded there).

# v6 = mixA (more real paired data): hospital repair 0.475/0.185 (v4/v5 was 0/0.42)
FROM hf.co/ricalanis/scrubdata-qwen3-4b-v6-q8:Q8_0
TEMPLATE """{{- if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{- range .Messages }}<|im_start|>{{ .Role }}
{{ .Content }}<|im_end|>
{{ end }}<|im_start|>assistant
"""
PARAMETER stop "<|im_end|>"
PARAMETER temperature 0