allenai/llama-3-tulu-v2-sft-subset

Recreating some subsets of Tulu 2 SFT Mix with Llama 3.1 405B completions from SambaNova

The following subsets (all generated with GPT-4):

• GPT4-Alpaca
• Open Orca
• Coda Alpaca

Not that the raw dataset includes all the tulu prompts, but some were empty via max_length issues on the API. We filtered with

is_any_content_empty = lambda obj_list: any(not item["content"] for item in obj_list)

Here's a basic script for generating this data:

import os

import jsonlines
import requests
import json 
from datasets import load_dataset

ds = load_dataset("allenai/tulu-v2-sft-mixture")
ds = ds["train"]

outfile = jsonlines.open('tulu2_subset_regenerations.jsonl', 'w')
# code_alpaca open_orca gpt4_alpaca oasst1
counter = 0
for row in ds:
    if row["dataset"] == "code_alpaca" or row["dataset"] == "open_orca" or row["dataset"] == "gpt4_alpaca":
        counter += 1
        if len(row["messages"]) > 2:
            print(row)
        dialogue = row["messages"]
        prompt = dialogue[0]["content"]
        max_tokens = 4000 - len(prompt.split())
        messages = []
        messages.append({"role": "system", "content": "You are a highly efficient assistant. You are to be as fair and accurate"})
        messages.append(
        {
            "role": "user",
            "content": prompt
        }
        )
        payload = {"messages": messages, "max_tokens": 4096, "stop": ["[INST", "[INST]", "[/INST]", "[/INST]"], "model": "llama3-405b", "stream": "true"}

        key = "your key"
        url = "https://7swuc05a91h3zixk.snova.ai/v1/chat/completions"

        headers = {"Authorization": f"Basic {key}", "Content-Type": "application/json"}
        post_response = requests.post(url, json=payload, headers=headers, stream=True)
        response_text = ""
        for line in post_response.iter_lines():
            if line.startswith(b"data: "):
                data_str = line.decode("utf-8")[6:]
                try:
                    line_json = json.loads(data_str)
                    if "choices" in line_json and "content" in line_json["choices"][0]["delta"]:
                        try:
                            response_text += line_json["choices"][0]["delta"]["content"]
                        except:
                            breakpoint()
                except json.JSONDecodeError as e:
                    pass
        out_dict = {}
        out_dict["dataset"] = row["dataset"]
        out_dict["id"] = row["id"]
        out_dict["regeneration_model"] = "llama3-405b"
        out_dict["messages"] = []
        out_dict["messages"].append({"role": "user", "content": prompt})
        out_dict["messages"].append({"role": "assistant", "content": response_text})
        outfile.write(out_dict)
        print(response_text)
        print(counter)