EuroPython2022/Scratchpad-w-BLOOM
14
1import gradio as gr2import requests3import os4 5##Bloom6API_URL = "https://api-inference.huggingface.co/models/bigscience/bloom"7HF_TOKEN = os.environ["HF_TOKEN"]8headers = {"Authorization": f"Bearer {HF_TOKEN}"}9 10def text_generate(prompt):11 print(f"Prompt is :{prompt}")12 p = prompt + " Solution: " 13 print(f"Final prompt is : {p}")14 json_ = {"inputs": p,15 "parameters":16 {17 "top_p": 0.9,18 "temperature": 1.1,19 "max_new_tokens": 250,20 "return_full_text": True21 }, "options": 22 {23 "use_cache": True,24 "wait_for_model":True25 },}26 response = requests.post(API_URL, headers=headers, json=json_)27 print(f"Response is : {response}")28 output = response.json()29 print(f"output is : {output}")30 output_tmp = output[0]['generated_text']31 print(f"output_tmp is: {output_tmp}")32 solution = output_tmp.split("\nQ:")[0]33 print(f"Final response after splits is: {solution}") 34 return solution 35 36demo = gr.Blocks()37 38with demo:39 gr.Markdown("<h1><center>Length generalization (LG) With BLOOM🌸 </center></h1>")40 gr.Markdown(41 """42 We will examine large language models ability to extrapolate to longer problems! \n43 Length generalization (LG) is important: Often, long examples are rare and intrinsically more difficult, yet are the ones we care more about. \n44 Recent paper [Exploring Length Generalization in Large Language Models](https://arxiv.org/pdf/2207.04901) found that using few-shot [scratchpad](https://arxiv.org/abs/2112.00114), a combo behind many strong LLM results (eg. #Minerva ) \n45 leads to **substantial improvements in length generalization!** \n46 In-context learning enables variable length pattern matching, producing solutions of correct lengths. \n47 This space is an attempt at inspecting this LLM behavior/capability in the new HuggingFace BigScienceW [Bloom](https://huggingface.co/bigscience/bloom) model. \n48 This Space is created by [Muhtasham Oblokulov](https://twitter.com/muhtasham9) for EuroPython 2022 Demo. \n49 This Space is work in progress, BLOOM doesn't support inference on long sequencess so you may try with shorter sequences. \n50 """51 )52 with gr.Row(): 53 input_prompt = gr.Textbox(value="Q:The coin is heads up.(1) Then Austin flips. Is the coin still heads up? Solution: Coin is initially heads up. (1) After Austin flips, coin turns to heads. Q: The coin is heads up. (2) Then Austin doesn't flip. (1) Then Kara flips. Is the coin still heads up?",54 label="Enter your examples zero-shot (few-shot is not supported due to API limit) followed by Query :")55 generated_txt = gr.Textbox(lines=10, label="Generated Solution:")56 57 b1 = gr.Button("Generate Text")58 b1.click(text_generate,inputs=[input_prompt], outputs=[generated_txt])59 60 with gr.Row(): 61 gr.Markdown("")62 63demo.launch(enable_queue=True, debug=True)