- 人工構造一批Prompt做種子。（Starting with a small seed set of human-written tasks）
- 每次把一些種子+后來生成的Prompt，放到Input里做few-shot examples，用LLM生成更多的Prompt；（Using the LLM to generate new instructions based on the seed tasks）
- 過濾掉質量太差的，修正能要的；（Filtering and refining the generated instructions）
- 把生成的所有Prompt，輸入LLM得到輸出結果；（Creating input-output instances for the new instructions）
- Input+Output，做LLM的訓練樣本（Using the generated dataset to fine-tune the LLM）
第2步，LLM生成：
import random from transformers import AutoTokenizer, AutoModelForCausalLM# Load a pre-trained language model model_name = "bigcode/starcoderbase-1b" tokenizer = AutoTokenizer.from_pretrained(model_name) model = AutoModelForCausalLM.from_pretrained(model_name)# Seed tasks (simplified for demonstration) seed_tasks = ["Write a function to calculate the factorial of a number.","Create a class to represent a bank account.","Implement a binary search algorithm." ]def generate_instruction(prompt):inputs = tokenizer(prompt, return_tensors="pt")outputs = model.generate(**inputs, max_new_tokens=50)return tokenizer.decode(outputs[0], skip_special_tokens=True)def self_instruct(num_iterations):generated_tasks = []for _ in range(num_iterations):# Sample existing taskssampled_tasks = random.sample(seed_tasks + generated_tasks, min(3, len(seed_tasks) + len(generated_tasks)))# Create a prompt for generating new instructionsprompt = "Generate a new programming task based on these examples:\n\n"prompt += "\n".join(sampled_tasks)prompt += "\n\nNew task:"# Generate a new instructionnew_task = generate_instruction(prompt)# In practice, you would filter and refine the generated task heregenerated_tasks.append(new_task)return generated_tasks# Run Self-Instruct new_tasks = self_instruct(5) for i, task in enumerate(new_tasks, 1):print(f"Task {i}: {task}")
第3步過濾：
人工定義一些規則，過濾掉太差的；（也可以用LLM來做裁判）
目的：確保質量和多樣性；
- Filter out instructions that are too short or too long
- Filter out instructions containing keywords unsuitable for language models (e.g. "image", "graph", "file", "plot")
- Filter out instructions starting with punctuation
- Filter out instructions starting with non-English characters
- Filter out instructions that have high ROUGE-L similarity (above 0.7) with any existing instruction in the task pool
)
