Rulga commited on
Commit
5a6f1a5
·
1 Parent(s): af6d411

Add XGLM 7.5B model details to app.py and settings.py for multilingual support

Browse files
Files changed (2) hide show
  1. app.py +22 -0
  2. config/settings.py +22 -0
app.py CHANGED
@@ -102,7 +102,29 @@ MODEL_DETAILS = {
102
  ],
103
  "documentation": "https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2"
104
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
105
  }
 
106
  # MODEL_DETAILS = {
107
  # "llama-7b": {
108
  # "full_name": "Meta Llama 2 7B Chat",
 
102
  ],
103
  "documentation": "https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2"
104
  },
105
+ "xglm-7.5b": {
106
+ "full_name": "Meta XGLM 7.5B",
107
+ "capabilities": [
108
+ "Specialized for multilingual generation",
109
+ "Support for 30+ languages",
110
+ "Strong cross-lingual transfer abilities",
111
+ "Consistent performance across diverse languages"
112
+ ],
113
+ "limitations": [
114
+ "Less instruction-tuned than dedicated chat models",
115
+ "May require more specific prompting",
116
+ "Not specifically optimized for legal domain",
117
+ "Slightly larger model requiring more GPU memory"
118
+ ],
119
+ "use_cases": [
120
+ "International legal assistance in native languages",
121
+ "Complex multilingual documentation",
122
+ "Serving clients from diverse linguistic backgrounds",
123
+ "Translation and summarization of legal concepts across languages"
124
+ ],
125
+ "documentation": "https://huggingface.co/facebook/xglm-7.5B"
126
  }
127
+ }
128
  # MODEL_DETAILS = {
129
  # "llama-7b": {
130
  # "full_name": "Meta Llama 2 7B Chat",
config/settings.py CHANGED
@@ -87,6 +87,28 @@ MODELS = {
87
  }
88
  }
89
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
90
  }
91
 
92
  # Default model
 
87
  }
88
  }
89
  },
90
+ "xglm-7.5b": {
91
+ "id": "facebook/xglm-7.5B",
92
+ "name": "XGLM 7.5B",
93
+ "description": "Meta's multilingual model designed for cross-lingual generation",
94
+ "type": "base",
95
+ "parameters": {
96
+ "max_length": 2048,
97
+ "temperature": 0.7,
98
+ "top_p": 0.9,
99
+ "repetition_penalty": 1.1,
100
+ },
101
+ "training": {
102
+ "base_model_path": "facebook/xglm-7.5B",
103
+ "fine_tuned_path": os.path.join(TRAINING_OUTPUT_DIR, "xglm-7.5b-tuned"),
104
+ "lora_config": {
105
+ "r": 16,
106
+ "lora_alpha": 32,
107
+ "lora_dropout": 0.05,
108
+ "target_modules": ["q_proj", "v_proj", "k_proj", "o_proj"]
109
+ }
110
+ }
111
+ }
112
  }
113
 
114
  # Default model