phi-2-slimorca

Model Card for Phi 2 SlimOrca

<!-- Provide a quick summary of what the model is/does. -->

Phi 2 finetuned on SlimOrca-Dedup. This model was trained with the goal of giving Phi 2 the ablity to generate the EOS token together with being capable of doing beam search. It can also follow custom system prompts as shown in the example below.

Model Details

How to Get Started with the Model

import torch
import transformers

model = transformers.AutoModelForCausalLM.from_pretrained(
    "miguelcarv/phi-2-slimorca",
    trust_remote_code=True,
).to('cuda')
tokenizer = transformers.AutoTokenizer.from_pretrained("microsoft/phi-2")


SYSTEM_PROMPT = "You are an AI assistant. You will be given a task. You must generate a short and concise answer."
input_text = f"""{SYSTEM_PROMPT}

Instruction: Give me the first 5 prime numbers and explain what prime numbers are. 
Output:"""

with torch.no_grad():
    outputs = model.generate(
        tokenizer(input_text, return_tensors="pt")['input_ids'].to('cuda'),
        max_length=1024,
        num_beams = 3,
        eos_token_id = tokenizer.eos_token_id
    )
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Details

• Trained for one epoch on SlimOrca-Dedup
• Learning rate: 1e-5
• Cosine learning rate decay to 0
• Optimizer: AdamW
• Batch size: 256
• Trained with mixed-precision bfloat16