| --- |
| library_name: transformers |
| tags: [] |
| --- |
| |
| # Model Card for Model ID |
|
|
| <!-- Provide a quick summary of what the model is/does. --> |
| Finetuned Phi-3-Small-8K-Instruct model on interaction data from Zooniverse. |
| It achieves 0.735 accuracy on character_interaction dataset (test split), which surpasses gpt-4o-2024-05-13's 0.699 accuracy on same split. |
| |
| |
| |
| ## Model Details |
| |
| ### Model Description |
| |
| <!-- Provide a longer summary of what this model is. --> |
| |
| |
| - **Developed by:** [Michael Xu] |
| - **Funded by:** [Andrew Piper] |
| - **Model type:** [Autoregressive language modeling] |
| - **Language(s) (NLP):** [English] |
| - **Finetuned from model:** [microsoft/Phi-3-small-8k-instruct] |
| |
| ### Model Sources [optional] |
| |
| <!-- Provide the basic links for the model. --> |
| |
| - **Repository:** [More Information Needed] |
| |
| |
| ## How to Get Started with the Model |
| |
| Use the code below to get started with the model. |
| |
| ``` |
| model = AutoModelForCausalLM.from_pretrained( |
| "ChunB1/Phi-3-interact", |
| torch_dtype="auto", |
| trust_remote_code=True, |
| attn_implementation="flash_attention_2", |
| ) |
| model.to("cuda") |
| tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3-small-8k-instruct") |
| |
| pipe = pipeline( |
| "text-generation", |
| model=model, |
| tokenizer=tokenizer, |
| ) |
| |
| # Using passage (two sentences), char1 and char2 to classify the interaction type (No, Associating, Thinking, Touching, Observing, Communicating) for one datapoint. |
| example = {'book_name': '2013_O_Connell,Carol_ItHappensintheDark_MY', |
| 'sentence_ID': 371, |
| 'passage': 'Her smile was just a flash, a taste of things to come. He shot her a glance to beg, Play nice.', |
| 'char1_COREF': 95, |
| 'char2_COREF': 448, |
| 'char1': 'He', |
| 'char2': 'her',} |
| |
| prompt_five_class_explained = """Communicating: char1 and char2 are engaged in some form of communication, such as speaking, writing, or signaling. |
| Associating: char1 and char2 are linked by a social or relational context, such as friendship, teamwork, or other associative bonds. |
| Observing: at least one character is observing or watching another one, without direct interaction. |
| Thinking: at least one character is thinking about or recalling memories of another one, without direct interaction. |
| Touching: char1 and char2 are engaged in physical touch or contact.""" |
| |
| prompt_base = "what kind of interaction between char1 and char2? Choose one of six options: No, Associating, Thinking, Touching, Observing, Communicating." |
|
|
| prompt = """Task Description: Classify the type of interaction between char1 and char2 in a given passage. There are six categories of interaction: |
|
|
| No interaction: Direct or indirect interaction does not occur between char1 and char2. Any imagination or assumption of interaction also counts as No. |
| """ + prompt_five_class_explained + prompt_base |
|
|
| prompt_suffix = "Only return the option and don't provide any extra information." |
| |
| prompt_full = f"passage: {example['passage']}, char1: {example['char1']}, char2: {example['char2']}, " + prompt + prompt_suffix |
| messages = [{"role": "user", "content": prompt_full}] |
|
|
| generation_args = { |
| "max_new_tokens": 15, |
| "return_full_text": False, |
| "do_sample": False, |
| } |
|
|
| # Label will be "Communicating" |
| print(pipe(messages, **generation_args)[0]["generated_text"].strip()) |
| ``` |
