Views
No views yet
"Influence of shallow infiltration on time-lapse ERT: Experience of advanced interpretation" Rémi Clément et al., C. R. Geoscience 341 (2009)
1from transformers import AutoModelForCausalLM, AutoTokenizer
2from peft import PeftModel
3
4# Chargement hiérarchique: Mistral → KIBALI Phase 1 → Expert 5
5base_model = AutoModelForCausalLM.from_pretrained(
6 "mistralai/Mistral-7B-Instruct-v0.2",
7 device_map="auto",
8 torch_dtype=torch.float16
9)
10
11# Appliquer KIBALI Phase 1
12kibali_base = PeftModel.from_pretrained(base_model, "BelikanM/kibali-instruct-7b-lora")
13
14# Appliquer Expert 5
15model = PeftModel.from_pretrained(kibali_base, "BelikanM/kibali-expert5-ert-advanced")
16tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2")
17
18# Inférence spécialisée ERT
19prompt = "[INST] Explique les artefacts de résistivité en ERT temporelle [/INST]"
20inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
21outputs = model.generate(**inputs, max_length=512)
22print(tokenizer.decode(outputs[0], skip_special_tokens=True))Mistral-7B (13GB)
↓
KIBALI Phase 1: Base scientifique (+161MB)
↓
├──> Expert 4: Multi-format .dat (+144MB)
└──> Expert 5: ERT avancé (+15MB) ← ICI