---
library_name: transformers
license: mit
datasets:
- HuggingFaceH4/ultrachat_200k
- Open-Orca/OpenOrca
language:
- en
---

# Model Card for Model ID

Phi-2-chat-v05 is a finetuned version of Phi-2 to increase the model's understanding of instructions and multi-turn conversations.
In essence: it now has a concept of shutting up after an answer is given - as opposed to just switching into random generator mode.

Finetuning used 25k records from the dataset `HuggingFaceH4/ultrachat_200k`


# Prompt format

```
<|system|>
You are a helpful assistant....
<|user|>
Why is the sky blue?
<|assistant|>
The sky appears blue because of a phenomenon called Rayleigh scattering. When sunlight enters the Earth's atmosphere [...]
<|user|>
Who was the phenomenon named after?
<|assistant|>
```

The system generates its output after the special token <|assistant|>. You need to have that token in the input for a reliable response.
Or you can use the tokenizer's chat_template, as shown below.

# How to use it?

Dependencies
```
pip install -u torch[cuda] transformers einops
```

Code for inference.


```python
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "WeeRobots/phi-2-chat-v05"

model = AutoModelForCausalLM.from_pretrained(model_id, device_map={"": 0}, trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=False, trust_remote_code=True)

payload = tokenizer.apply_chat_template([
    { 'role': 'system', 'content': '''You are a state machine. The user will add state slot values and you'll keep track of them.''' },
    { 'role': 'user', 'content': '''Place 15 into slot apple''' },
    { 'role': 'assistant', 'content': '''Roger that.''' },
    { 'role': 'user', 'content': '''Bananas slot should be 20''' },
    { 'role': 'assistant', 'content': '''Certainly''' },
    { 'role': 'user', 'content': '''What is the value of Apple + Banana?''' },
], tokenize=False, add_generation_prompt=True,)
device = "cuda"
model_input = tokenizer(payload, return_tensors="pt").to(device)
with torch.no_grad():
  # IMPORTANT: always set the eos_token_id in this call. the model is trained to emit the eos_token the right time
  # but it might continue generating irrelevant text. this way the model will stop at the right place
  model_response = model.generate(**model_input, max_new_tokens=512, eos_token_id=tokenizer.eos_token_id, )
  print(tokenizer.decode(model_result[0], skip_special_tokens=False))
```

# Non production quality
Be aware that this model tuning wasn't thoroughly tested, and isn't meant to be used in production, only for experimentation or hobby projects.