README.md

---
license: mit
library_name: transformers
base_model: microsoft/phi-2
datasets:
- teknium/OpenHermes-2.5
- ContextualAI/ultrafeedback_clair_32k
pipeline_tag: text-generation
model-index:
- name: phi-2-instruct-apo
  results:
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: IFEval (0-Shot)
      type: HuggingFaceH4/ifeval
      args:
        num_few_shot: 0
    metrics:
    - type: inst_level_strict_acc and prompt_level_strict_acc
      value: 31.46
      name: strict accuracy
    source:
      url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=rasyosef/phi-2-instruct-apo
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: BBH (3-Shot)
      type: BBH
      args:
        num_few_shot: 3
    metrics:
    - type: acc_norm
      value: 21.67
      name: normalized accuracy
    source:
      url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=rasyosef/phi-2-instruct-apo
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: MATH Lvl 5 (4-Shot)
      type: hendrycks/competition_math
      args:
        num_few_shot: 4
    metrics:
    - type: exact_match
      value: 0.0
      name: exact match
    source:
      url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=rasyosef/phi-2-instruct-apo
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: GPQA (0-shot)
      type: Idavidrein/gpqa
      args:
        num_few_shot: 0
    metrics:
    - type: acc_norm
      value: 2.68
      name: acc_norm
    source:
      url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=rasyosef/phi-2-instruct-apo
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: MuSR (0-shot)
      type: TAUR-Lab/MuSR
      args:
        num_few_shot: 0
    metrics:
    - type: acc_norm
      value: 3.61
      name: acc_norm
    source:
      url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=rasyosef/phi-2-instruct-apo
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: MMLU-PRO (5-shot)
      type: TIGER-Lab/MMLU-Pro
      config: main
      split: test
      args:
        num_few_shot: 5
    metrics:
    - type: acc
      value: 12.83
      name: accuracy
    source:
      url: https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard?query=rasyosef/phi-2-instruct-apo
      name: Open LLM Leaderboard
---

# Phi-2-Instruct-APO

This is a finetuned version of Microsoft's 2.7B parameter [phi-2](https://huggingface.co/microsoft/phi-2) transfromer model that has underwent a post-training process that incorporates both **supervised fine-tuning** and **anchored preference optimization** for instruction following. I used the [trl](https://huggingface.co/docs/trl/en/index) library and a single **A100 40GB** GPU during both the SFT and APO steps.

- Supervised Fine-Tuning
  - SFT Model: [phi-2-sft](https://huggingface.co/rasyosef/phi-2-sft-openhermes-128k-v2)
  - Used 128,000 instruction, response pairs from the [teknium/OpenHermes-2.5](https://huggingface.co/datasets/teknium/OpenHermes-2.5) dataset

- Anchored Preference Optimization (APO)
  - LoRA Adapter: [phi-2-apo](https://huggingface.co/rasyosef/phi-2-apo)
  - Used 10,000 preference pairs from the [ContextualAI/ultrafeedback_clair_32k](https://huggingface.co/datasets/ContextualAI/ultrafeedback_clair_32k) dataset

## How to use
### Chat Format

Given the nature of the training data, the phi-2 instruct model is best suited for prompts using the chat format as follows. 
You can provide the prompt as a question with a generic template as follows:
```markdown
<|im_start|>system
You are a helpful assistant.<|im_end|>
<|im_start|>user
Question?<|im_end|>
<|im_start|>assistant
```

For example:
```markdown
<|im_start|>system
You are a helpful assistant.<|im_end|>
<|im_start|>user
How to explain Internet for a medieval knight?<|im_end|>
<|im_start|>assistant
```
where the model generates the text after `<|im_start|>assistant` .

### Sample inference code

This code snippets show how to get quickly started with running the model on a GPU:

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

torch.random.manual_seed(0) 

model_id = "rasyosef/phi-2-instruct-apo"
model = AutoModelForCausalLM.from_pretrained( 
    model_id,  
    device_map="cuda",  
    torch_dtype="auto" 
) 

tokenizer = AutoTokenizer.from_pretrained(model_id) 

messages = [ 
    {"role": "system", "content": "You are a helpful AI assistant."}, 
    {"role": "user", "content": "Can you provide ways to eat combinations of bananas and dragonfruits?"}, 
    {"role": "assistant", "content": "Sure! Here are some ways to eat bananas and dragonfruits together: 1. Banana and dragonfruit smoothie: Blend bananas and dragonfruits together with some milk and honey. 2. Banana and dragonfruit salad: Mix sliced bananas and dragonfruits together with some lemon juice and honey."}, 
    {"role": "user", "content": "What about solving an 2x + 3 = 7 equation?"}, 
] 

pipe = pipeline( 
    "text-generation", 
    model=model, 
    tokenizer=tokenizer, 
) 

generation_args = { 
    "max_new_tokens": 256, 
    "return_full_text": False, 
    "temperature": 0.0, 
    "do_sample": False, 
} 

output = pipe(messages, **generation_args) 
print(output[0]['generated_text'])  
```

Note: If you want to use flash attention, call _AutoModelForCausalLM.from_pretrained()_ with _attn_implementation="flash_attention_2"_


## Benchmarks

These benchmarks were run using EleutherAI's [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness)

- **IFEval (Instruction Following Evaluation)**: IFEval is a fairly interesting dataset that tests the capability of models to clearly follow explicit instructions, such as “include keyword x” or “use format y”. The models are tested on their ability to strictly follow formatting instructions rather than the actual contents generated, allowing strict and rigorous metrics to be used.
- **GSM8k (5-shot, flexible-extract)**: diverse grade school math word problems to measure a model's ability to solve multi-step mathematical reasoning problems.
- **MMLU (5-shot)** - a test to measure a text model's multitask accuracy. The test covers 57 tasks including elementary mathematics, US history, computer science, law, and more.
- **TruthfulQA** - a test to measure a model's propensity to reproduce falsehoods commonly found online. Note: TruthfulQA is technically a 6-shot task in the Harness because each example is prepended with 6 Q/A pairs, even in the 0-shot setting.
- **Winogrande (5-shot)** - an adversarial and difficult Winograd benchmark at scale, for commonsense reasoning.

|Model |[Phi-2-Instruct-APO](https://huggingface.co/rasyosef/phi-2-instruct-apo)|[Phi-2](https://huggingface.co/microsoft/phi-2)|
|:-----|:-------------------------------------------------------------------------|:----------------------------------------------|
|Size (# params)|2.7B|2.7B|
|IFEval|**34.48**|26.53|
|GSM8K|52.16|**56.44**|
|MMLU|44.88|**56.70**|
|TruthfulQA|**49.44**|44.48|
|Winogrande|**75.61**|73.72|

## Code

https://github.com/rasyosef/phi-2-apo
# [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard)
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_rasyosef__phi-2-instruct-apo)

|      Metric       |Value|
|-------------------|----:|
|Avg.               |12.04|
|IFEval (0-Shot)    |31.46|
|BBH (3-Shot)       |21.67|
|MATH Lvl 5 (4-Shot)| 0.00|
|GPQA (0-shot)      | 2.68|
|MuSR (0-shot)      | 3.61|
|MMLU-PRO (5-shot)  |12.83|