Create README.md

README.md

@@ -0,0 +1,67 @@
+```
+#!/usr/bin/env python
+import torch
+from transformers import (
+    AutoConfig,
+    AutoTokenizer,
+    AutoModelForCausalLM,
+    LlamaForSequenceClassification,
+)
+# install torch, transformers, accelerate
+
+def main():
+    # Define the input and output repository names.
+    input_model_id = "meta-llama/Meta-Llama-3-70B-Instruct"
+    split_2 = input_model_id.split("/")[1]
+    output_model_id = f"baseten/example-{split_2}ForSequenceClassification"  
+
+    # Load the original configuration.
+    # (If needed, add trust_remote_code=True for custom implementations.)
+    config = AutoConfig.from_pretrained(input_model_id)
+    
+    # Update the config for a sequence classification task with 10 labels.
+    num_labels = 30
+    config.num_labels = num_labels
+    config.id2label = {i: f"token activation {i}" for i in range(num_labels)}
+    config.label2id = {f"token activation {i}": i for i in range(num_labels)}
+
+    # Download the tokenizer from the original model.
+    tokenizer = AutoTokenizer.from_pretrained(input_model_id)
+
+    # Load the original causal LM model.
+    lm_model = AutoModelForCausalLM.from_pretrained(input_model_id, config=config, device_map="auto", low_cpu_mem_usage=True)
+    config.architectures = ["LlamaForSequenceClassification"]
+    del lm_model.model
+    print("loaded lm model")
+    # Initialize the sequence classification model.
+    # NOTE: We are using the built-in LlamaForSequenceClassification,
+    # which uses a `.score` attribute as the output head.
+    seq_cls_model = LlamaForSequenceClassification.from_pretrained(input_model_id, config=config, device_map="auto", low_cpu_mem_usage=True)
+
+    # --- Initialize the Classification Head ---
+    # Here we re-use the first 10 rows from the original LM head
+    # (i.e. rows 0 to 9) to initialize the new classification head.
+    with torch.no_grad():
+        # lm_model.lm_head.weight has shape [vocab_size, hidden_size]
+        # We take the first 10 rows to form a [10, hidden_size] weight matrix.
+        seq_cls_model.score.weight.copy_(lm_model.lm_head.weight.data[:num_labels, :])
+        if lm_model.lm_head.bias is not None:
+            seq_cls_model.score.bias.copy_(lm_model.lm_head.bias.data[:num_labels])
+
+    # Optionally, save the new model locally.
+    # save_directory = f"./{output_model_id.replace('/','_')}"
+    # seq_cls_model.save_pretrained(save_directory)
+    # tokenizer.save_pretrained(save_directory)
+
+    # Push the new model and tokenizer to the Hub.
+    # (Ensure you are authenticated with Hugging Face Hub via `huggingface-cli login`.)
+    tokenizer.push_to_hub(output_model_id)
+    seq_cls_model.push_to_hub(output_model_id)
+    
+
+    print(f"New model pushed to the Hub: {output_model_id}")
+
+if __name__ == "__main__":
+    main()
+
+```