turtle170/Axiom-30

---
language:
- en
license: apache-2.0
task_categories:
- text-generation
- question-answering
tags:
- mathematics
- reasoning
- chain-of-thought
- instruction-tuning
- synthetic-data
- axiomatic-reasoning
- filtered
- extreme-quality
- ChatML
pretty_name: Axiom-30
size_categories:
- 10M<n<100M
---

**Dataset Details:**

Axiom-30 is a dataset containing 30 million samples, all above 500 characters, adapted from the OpenMath, OpenThought, and UltraChat datasets.

The dataset is separated into 3 sections:

The math section,

The thought section,

and finally the chat section.

**Important Notes:**
The uncompressed .jsonl file from the Gzip file will be around 125GB.

This datasethas three distinct formatting. The pytho code for sorting them is:

def formatting_prompts_func(examples):
instructions = examples["problem"] if "problem" in examples else examples["instruction"]
# If it's UltraChat, we might need to pull from 'messages'
responses = examples["generated_solution"] if "generated_solution" in examples else examples["output"]

texts = []
for instr, resp in zip(instructions, responses):
# This converts everything to ChatML on-the-fly!
text = f"<|im_start|>user\n{instr}<|im_end|>\n<|im_start|>assistant\n{resp}<|im_end|>"
texts.append(text)
return { "text": texts }