ajvikram's picture
Create Modelfile or ollama
331c480 verified
# FROM ... (Keep your existing FROM line)
#FROM hf.co/ajvikram/Qwen2.5-Coder-3B-Network-Eng-GGUF
FROM hf.co/ajvikram/network-compliance-auditor-grpo
# --- 1. STOP TOKENS (Keep these!) ---
PARAMETER stop "<|im_end|>"
PARAMETER stop "</answer>"
PARAMETER stop "User:"
PARAMETER stop "<|fim_prefix|>"
PARAMETER stop "<|fim_middle|>"
PARAMETER stop "<|fim_suffix|>"
PARAMETER stop "<|repo_name|>"
# --- 2. CRITICAL FIX: ANTI-LOOPING ---
# repeat_penalty 1.15 tells the model: "If you just said this token, don't say it again."
# This kills the "😦 😦 😦" loop instantly.
PARAMETER repeat_penalty 1.15
# repeat_last_n 64 tells it to look back at the last 64 tokens to check for repeats.
PARAMETER repeat_last_n 64
# --- 3. TEMPLATE (Keep this exactly as before) ---
TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
{{ end }}<|im_start|>assistant
<think>
{{ .Response }}"""
# --- 4. SYSTEM (Keep this exactly as before) ---
SYSTEM """ You are an intelligent Network Compliance Auditor. You must output your internal reasoning trace within <reasoning> tags, followed by a final JSON object containing the verdict (PASS/FAIL), the specific control ID, and the evidence line. Dont repeat and provide only one answer."""