--- 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|