Upload model_info.json with huggingface_hub

model_info.json

@@ -0,0 +1,40 @@
+{
+  "model_name": "bnb_8bit_conservative_stem_optimized",
+  "source_model": "luca-deandrea/MNLP_M3_mcqa_model",
+  "quantization_method": "BitsAndBytes_8bit",
+  "timestamp": "2025-06-10 23:38:18",
+  "processing_time_seconds": 62.69236969947815,
+  "quantization_config": {
+    "load_in_8bit": true,
+    "llm_int8_threshold": 8.0,
+    "llm_int8_skip_modules": [
+      "lm_head"
+    ],
+    "llm_int8_enable_fp32_cpu_offload": false,
+    "llm_int8_has_fp16_weight": false
+  },
+  "model_stats": {
+    "size_mb": 732.8867292404175,
+    "validation_score": 0.4,
+    "compression_ratio": "~2x compared to FP16"
+  },
+  "stem_optimizations": [
+    "conservative_8bit_threshold_8.0",
+    "lm_head_layer_kept_in_fp16",
+    "deterministic_generation_config",
+    "optimized_for_stem_reasoning",
+    "cache_enabled_for_performance"
+  ],
+  "usage_instructions": {
+    "library": "transformers",
+    "loading": "AutoModelForCausalLM.from_pretrained('thresh10')",
+    "recommended_for": "MNLP_STEM_mcqa_evals",
+    "notes": "Model ready for lighteval evaluation"
+  },
+  "performance_characteristics": {
+    "accuracy": "High (conservative quantization)",
+    "speed": "Good (8-bit quantization)",
+    "memory": "Reduced (~50% of original)",
+    "best_for": "STEM reasoning tasks requiring high accuracy"
+  }
+}