22 lines
1.3 KiB
Markdown
22 lines
1.3 KiB
Markdown
---
|
|
library_name: transformers
|
|
pipeline_tag: text-generation
|
|
base_model:
|
|
- Qwen/Qwen3-8B
|
|
---
|
|
|
|
# DKatiyar-fixed
|
|
|
|
Fixed version of [yunmorning/broken-model](https://huggingface.co/yunmorning/broken-model).
|
|
|
|
## Changes Made
|
|
|
|
### 1. Added missing `chat_template` to `tokenizer_config.json`
|
|
|
|
The original repo had no `chat_template` field. Without it, the serving engine cannot format role-based messages (system, user, assistant) with the special tokens the model expects (`<|im_start|>`, `<|im_end|>`). Raw text is sent instead, producing garbled output because the model was trained on the ChatML format.
|
|
|
|
The fix: copied the `chat_template` from the official [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) repository. This template handles system/user/assistant role formatting, tool calls (`<tool_call>`/`</tool_call>`), thinking mode (`<think>`/`</think>`), and the `enable_thinking` toggle.
|
|
|
|
### 2. Corrected `base_model` in README metadata
|
|
|
|
The original README listed `meta-llama/Meta-Llama-3.1-8B` as the base model. The actual architecture is Qwen3: `config.json` specifies `Qwen3ForCausalLM` with 36 layers, 4096 hidden_size, 32 attention heads, 8 KV heads, and vocab_size 151936. The tokenizer is `Qwen2Tokenizer`. These all match [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B), not Meta-Llama. |