25 lines
782 B
Plaintext
25 lines
782 B
Plaintext
# FRANKENSTALLM 3B v2 Korean ORPO — Ollama Modelfile (Q4_K_M)
|
|
# Usage:
|
|
# ollama create frankenstallm-3b-v2:Q4_K_M -f Modelfile.3b-v2-Q4
|
|
# ollama run frankenstallm-3b-v2:Q4_K_M
|
|
#
|
|
# Sampling config: ORPO eval grid best (t0.7_rep1.2)
|
|
# 3-gram rep=0%, EOS=100%, avg_tokens=189.2
|
|
|
|
FROM ./outputs/gguf/frankenstallm-3b-v2-Q4_K_M.gguf
|
|
|
|
# --- Sampling Config (ORPO eval grid + Ollama 검증) ---
|
|
PARAMETER temperature 0.7
|
|
PARAMETER repeat_penalty 1.2
|
|
PARAMETER top_p 0.9
|
|
PARAMETER top_k 50
|
|
PARAMETER num_predict 512
|
|
PARAMETER num_ctx 4096
|
|
PARAMETER stop "</s>"
|
|
|
|
# --- System Prompt ---
|
|
SYSTEM """당신은 FRANKENSTALLM, 한국어에 특화된 AI 어시스턴트입니다. 정확하고 자연스러운 한국어로 답변해주세요."""
|
|
|
|
# --- License ---
|
|
LICENSE """Apache-2.0"""
|