A LLM trained only on data from certain time periods to reduce modern bias
Python
Add files via upload
Hayk Grigorian2592f0566716
diff --git a/london_1800_1875_v2mini_eval1/test_v2mini_eval2.py b/london_1800_1875_v2mini_eval1/test_v2mini_eval2.pynew file mode 100644index 0000000..40f6d3b--- /dev/null+++ b/london_1800_1875_v2mini_eval1/test_v2mini_eval2.py@@ -0,0 +1,83 @@+# run script for TimeCapsuleLLM 1800-1875 London v2mini-eval2+import torch+from transformers import LlamaForCausalLM, AutoTokenizer++# config+MODEL_DIR = "put your folder path here"+TOKENIZER_DIR = "put your path to tokenizer"+++# device setup+if torch.cuda.is_available():++ device = "cuda"+ dtype = torch.float16+ print(f"[Device] Using GPU: {torch.cuda.get_device_name(0)}")+else:+ device = "cpu"+ dtype = torch.float32++ print("[Device] CUDA not available — using CPU")+++# load tokenizer+print(f"[Tokenizer] Loading tokenizer from: {TOKENIZER_DIR}")+tokenizer = AutoTokenizer.from_pretrained(++ TOKENIZER_DIR,++ local_files_only=True+)+++# load model+print(f"[Model] Loading model from: {MODEL_DIR}")++model = LlamaForCausalLM.from_pretrained(+ MODEL_DIR,+ dtype=dtype,+ device_map="auto",+)+model.eval()++# generation function+def infer(prompt, max_new_tokens=200):+ print("\n==============================")+ print("PROMPT:")++ print(prompt)+ print("==============================")++ inputs = tokenizer(+ prompt,++ return_tensors="pt"+++ )+++ inputs.pop("token_type_ids", None)+ inputs = inputs.to(model.device)++ with torch.no_grad():+ output = model.generate(+ **inputs,+ max_new_tokens=max_new_tokens,++ temperature=0.7,+ top_p=0.9,++ do_sample=True,+ eos_token_id=tokenizer.eos_token_id,+ )++ text = tokenizer.decode(output[0], skip_special_tokens=True)+ print(text)++ print("==============================\n")+++# test prompts+infer("Put your prompt here")+infer("Theres no QA training, so questions dont work")