Luigi commited on
Commit
a38e521
·
1 Parent(s): d52338f

Remove Qwen2.5 3B - keep only 1.5B as extraction model

Browse files

Qwen2.5 3B showed poor performance:
- Misclassified action_items as key_points (0 action items extracted)
- 14.3% duplicate rate
- Slower (116s vs 80s)
- No quality improvement over 1.5B

Qwen2.5 1.5B is the optimal choice:
- Perfect categorization
- Zero duplicates
- Faster processing
- Proven 100% extraction success

Files changed (1) hide show
  1. app.py +0 -16
app.py CHANGED
@@ -689,22 +689,6 @@ EXTRACTION_MODELS = {
689
  "repeat_penalty": 1.0,
690
  },
691
  },
692
- "qwen2.5_3b": {
693
- "name": "Qwen2.5 3B (128K Context)",
694
- "repo_id": "Qwen/Qwen2.5-3B-Instruct-GGUF",
695
- "filename": "*q4_k_m.gguf",
696
- "max_context": 131072,
697
- "default_n_ctx": 4096,
698
- "params_size": "3B",
699
- "supports_reasoning": False,
700
- "supports_toggle": False,
701
- "inference_settings": {
702
- "temperature": 0.2,
703
- "top_p": 0.9,
704
- "top_k": 30,
705
- "repeat_penalty": 1.0,
706
- },
707
- },
708
  }
709
 
710
  DEFAULT_EXTRACTION_MODEL = "qwen2.5_1.5b"
 
689
  "repeat_penalty": 1.0,
690
  },
691
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
692
  }
693
 
694
  DEFAULT_EXTRACTION_MODEL = "qwen2.5_1.5b"