robnav/fai_bm_fix2

Minimal Fix for yunmorning/broken-model

This repository provides a structural fix for the yunmorning/broken-model, addressing critical metadata mismatches that prevent standard inference.

🛠 Problem

The original model failed to function with the FriendliAI v1/chat/completions endpoint. When queried via a Python wrapper, it triggered the following error:

no configured chat prompt template; use non-chat endpoint instead

🔍 Analysis

• Architecture Mismatch: While the human-maintained README originally cited meta-llama/Meta-Llama-3.1-8B, the machine-maintained config.json and tokenizer_config.json correctly identify a Qwen3 architecture.
• Base Model Identification: Further investigation confirms the true base model is Qwen/Qwen3-8B.
• Root Cause: The local tokenizer_config.json was missing the chat_template key, which is required for the Chat Completions API to format messages into the model's expected internal prompt structure.

🩹 Fix

To resolve these issues while maintaining a clean repository structure:

1. Metadata Alignment: Updated README.md YAML front matter to reflect the correct base_model (Qwen3).
2. Template Implementation: Decided to separate the large (~4KB) Jinja template into a standalone file for better maintainability and readability.
3. New File Added: chat_template.jinja. This file contains the pretty-printed Jinja2 logic sourced from the official Qwen3 reference.

🚀 Usage

With chat_template.jinja present in the repository root, modern inference engines will automatically detect and use the correct formatting for Chat completions.

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("robnav/fai_bm_fix2")
# The template is now automatically loaded from chat_template.jinja