Updated README.me

created proper readme

README.md

@@ -1,3 +1,148 @@
----
-license: mit
----
+---
+license: mit
+---
+
+#datasets
+
+- Yale-LILY/FOLIO
+- yuan-yang/MALLS-v0
+- apergo-ai/text2log (1661 records)
+
+how to load
+
+```
+device = "cuda"
+model_name_or_path = "microsoft/Phi-3-mini-4k-instruct"
+
+model = transformers.AutoModelForCausalLM.from_pretrained(
+    model_name_or_path, torch_dtype=torch.bfloat16, device_map=device)
+
+reft_model = pyreft.ReftModel.load(
+    "LaferriereJC/Phi-3-mini-4k-instruct-FOL-pyreft", model
+)
+```
+
+how to use
+```
+!git clone https://huggingface.co/LaferriereJC/Phi-3-mini-4k-instruct-FOL-pyreft
+from transformers import AutoModelForCausalLM
+import torch
+import pyreft
+import os
+import transformers
+
+device = 'cuda'
+model_name_or_path = "microsoft/Phi-3-mini-4k-instruct"
+
+attn_implementation = "eager"
+torch_dtype = torch.float16
+#"microsoft/Phi-3-mini-4k-instruct"
+
+model = transformers.AutoModelForCausalLM.from_pretrained(
+    model_name_or_path, torch_dtype=torch.bfloat16, device_map=device,trust_remote_code=True)
+
+
+# Define the PyReFT configuration
+layers = range(model.config.num_hidden_layers)
+representations = [{
+    "component": f"model.layers[{l}].output",
+    "intervention": pyreft.LoreftIntervention(
+        embed_dim=model.config.hidden_size, 
+        low_rank_dimension=16
+    )
+} for l in layers]
+
+reft_config = pyreft.ReftConfig(representations=representations)
+
+# Initialize the PyReFT model
+reft_model = pyreft.get_reft_model(model, reft_config)
+
+# Load the saved PyReFT model
+local_directory = "./Phi-3-mini-4k-instruct-FOL-pyreft"
+interventions = {}
+for l in layers:
+    component = f"model.layers[{l}].output"
+    file_path = os.path.join(local_directory, f"intkey_comp.{component}.unit.pos.nunit.1#0.bin")
+    if os.path.exists(file_path):
+        with open(file_path, "rb") as f:
+            adjusted_key = f"comp.{component}.unit.pos.nunit.1#0"
+            interventions[adjusted_key] = torch.load(f)
+
+# Apply the loaded weights to the model
+for component, state_dict in interventions.items():
+    if component in reft_model.interventions:
+        reft_model.interventions[component][0].load_state_dict(state_dict)
+    else:
+        print(f"Key mismatch: {component} not found in reft_model.interventions")
+
+# Set the device to CUDA
+reft_model.set_device("cuda")
+
+# Verify the model
+reft_model.print_trainable_parameters()
+
+#model.half()
+# get tokenizer
+tokenizer = transformers.AutoTokenizer.from_pretrained(
+    model_name_or_path, model_max_length=216,
+    padding_side="right", use_fast=True,
+    attn_implementation=attn_implementation
+    #, add_eos_token=True, add_bos_token=True
+)
+
+tokenizer.pad_token = tokenizer.eos_token
+
+# position info about the interventions
+share_weights = True # whether the prefix and suffix interventions sharing weights.
+positions="f3+l3"    # the intervening positions of prefix tokens (f[irst]1) and suffix tokens (l[ast]1).
+first_n, last_n = pyreft.parse_positions(positions)
+
+terminators = [
+    tokenizer.eos_token_id,
+]
+
+prompt_no_input_template = """\n<|user|>:%s</s>\n<|assistant|>:"""
+
+test_instruction = f"""tell me something I don't know"""
+# tokenize and prepare the input
+prompt = prompt_no_input_template % test_instruction
+prompt = tokenizer(prompt, return_tensors="pt").to(device)
+
+unit_locations = torch.IntTensor([pyreft.get_intervention_locations(
+    last_position=prompt["input_ids"].shape[-1], 
+    first_n=first_n, 
+    last_n=last_n,
+    pad_mode="last",
+    num_interventions=len(reft_config.representations),
+    share_weights=share_weights
+)]).permute(1, 0, 2).tolist()
+
+_, reft_response = reft_model.generate(
+    prompt, unit_locations={"sources->base": (None, unit_locations)},
+    intervene_on_prompt=True, max_new_tokens=216, do_sample=True, top_k=50,temperature=0.7,
+    eos_token_id=terminators, early_stopping=True
+)
+print(tokenizer.decode(reft_response[0], skip_special_tokens=True))
+
+
+```
+
+response
+```
+:tell me something I don't know</s> :exists x1.(_thing(x1) & _donknow(x1))
+```
+
+training settings
+```
+    per_device_train_batch_size=6,
+    logging_steps=1,
+    optim='paged_lion_8bit',
+    gradient_checkpointing_kwargs={"use_reentrant": False},
+    learning_rate=0.0003,
+    warmup_ratio=.1,
+    adam_beta2=0.95,
+    adam_epsilon=0.00001,
+    save_strategy='epoch',
+    max_grad_norm=1.0,
+    lr_scheduler_type='cosine',
+```