backend/llm/models/Modelfile

# VidyaBot Fine-Tuned Mistral Modelfile
#
# This file is used AFTER running modal_convert_gguf.py,
# which downloads and converts the fine-tuned model to GGUF format.
#
# Usage (once GGUF file is present):
#   ollama create mistral-vidyabot -f backend/llm/models/Modelfile
#
# The GGUF path is set by modal_convert_gguf.py automatically.
# If you need to run manually, update the FROM path below:
#
# FROM ./mistral-vidyabot.Q4_K_M.gguf

FROM ./mistral-vidyabot.Q4_K_M.gguf

# Context window — matches VidyaBot's pruning pipeline budget (512 tokens context + 256 output)
PARAMETER num_ctx 2048
PARAMETER num_predict 256
PARAMETER temperature 0.7
PARAMETER top_p 0.9
PARAMETER repeat_penalty 1.1
PARAMETER stop "<s>"
PARAMETER stop "</s>"
PARAMETER stop "[INST]"
PARAMETER stop "[/INST]"

# System prompt — baked into every query for consistent tutoring behaviour
SYSTEM """You are VidyaBot, an expert AI tutor for Indian school students studying NCERT curriculum (Classes 6-12). Your role is to help students understand concepts clearly.

Guidelines:
- Give direct, precise answers in 2-4 sentences
- Use the exact terminology found in NCERT textbooks
- For science questions: mention the relevant process, formula, or law
- For math questions: show the key formula or method
- If the question is in Hindi, answer in Hindi
- Keep answers concise — students need clarity, not essays"""