sangwoonoel/fixed-model

Overview

This model, sangwoonoel/fixed-model, is an 8 billion parameter Qwen3 architecture, specifically identified as Qwen3-8B through inspection of its weight map and configuration files. It features a 32K token context length.

Key Fixes and Characteristics

• Root Cause: The original yunmorning/broken-model was missing the chat_template field in its tokenizer_config.json.
• Solution: The chat_template field, essential for the /chat/completions API, has been added, adopting the official ChatML template expected by Qwen3 models.
• Verification: The model's identity as Qwen3-8B was confirmed by analyzing self_attn.q_norm / self_attn.k_norm layers, 36 hidden layers, and a vocab_size of 151936. Its config.json and generation_config.json are identical to the official Qwen/Qwen3-8B.
• Metadata Note: While the HuggingFace metadata incorrectly lists base_model: meta-llama/Meta-Llama-3.1-8B, this is a display-only inaccuracy and does not affect the model's functionality or the fix implemented.

Good For

• Applications requiring reliable chat/completions API functionality with Qwen3 models.
• Developers needing a Qwen3-8B model with correctly configured ChatML message formatting.
• Use cases where adherence to the <|im_start|> / <|im_end|> token structure is critical for chat interactions.