Phi-2-super (SFT + cDPO)

Base Model: microsoft/phi-2

image/png

How to run inference:

import transformers
import torch

if __name__ == "__main__":
  model_name = "abacaj/phi-2-super"
  tokenizer = transformers.AutoTokenizer.from_pretrained(model_name)
  
  model = (
      transformers.AutoModelForCausalLM.from_pretrained(
          model_name,
      )
      .to("cuda:0")
      .eval()
  )
  
  messages = [
      {"role": "user", "content": "Hello, who are you?"}
  ]
  inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
  input_ids_cutoff = inputs.size(dim=1)
  
  with torch.no_grad():
      generated_ids = model.generate(
          input_ids=inputs,
          use_cache=True,
          max_new_tokens=512,
          temperature=0.2,
          top_p=0.95,
          do_sample=True,
          eos_token_id=tokenizer.eos_token_id,
          pad_token_id=tokenizer.pad_token_id,
      )
  
  completion = tokenizer.decode(
      generated_ids[0][input_ids_cutoff:],
      skip_special_tokens=True,
  )
  
  print(completion)

Chat template

The model uses the same chat template as found in Mistral instruct models:

text = "<|endoftext|>[INST] What is your favourite condiment? [/INST]"
"Well, I'm quite partial to a good squeeze of fresh lemon juice. It adds just the right amount of zesty flavour to whatever I'm cooking up in the kitchen!<|endoftext|> "
"[INST] Do you have mayonnaise recipes? [/INST]"

You don't need to do it manually if you use the HF transformers tokenizer:

  messages = [
      {"role": "user", "content": "Hello, who are you?"},
      {"role": "assistant": "content": "I am ..."}
  ]
  inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)

MT-bench / heval

image/png image/png

Downloads last month
168
Safetensors
Model size
2.78B params
Tensor type
BF16
Β·
Inference Providers NEW
This model is not currently available via any of the supported Inference Providers.

Model tree for abacaj/phi-2-super

Finetunes
1 model
Merges
3 models
Quantizations
2 models

Spaces using abacaj/phi-2-super 3

Evaluation results

  • prompt_level_loose_acc on Instruction Following Eval
    LightEval
    0.272