grenishrai/math-sft-dataset

Dataset Card for Mori SFT Dataset

Mixed instruction data for supervised fine-tuning of the Mori small-model series. Math reasoning is the main push; a general-instruction slice is kept so compact models (around 1B–2B, including SmolLM2-1.7B) do not drop everyday instruction following.

Each row is a 3-turn conversation: a fixed system prompt, one user query, and one assistant response.

Dataset Details

Dataset Description

A single shuffled train split in Hugging Face chat messages format. Sources are GSM8K (cleaned), MetaMathQA-40K, and a 40k single-turn slice of OpenHermes-2.5. Rows are rewritten from {question, answer} into messages and shuffled with seed 42.

• Languages: English
• License: mixed — MIT (GSM8K) and Apache 2.0 (MetaMathQA, OpenHermes-2.5). Review each upstream license before redistribution.
• Format: JSONL, OpenAI-style messages
• Size: 87,314 training examples, one split (train)
• Held-out test split: none

Composition

Source	Examples	Focus
GSM8K (cleaned)	7,314	Grade-school math word problems
MetaMathQA-40K	40,000	Math reasoning, question/answer augmented
OpenHermes-2.5 (40k slice)	40,000	General instruction following
Total	87,314	Math-heavy mix with a general-knowledge anchor

Math examples (GSM8K + MetaMathQA) are the majority. The OpenHermes slice is the general-knowledge anchor against catastrophic forgetting.

Dataset Sources

• GSM8K: filtered from the original train split (7,473 → 7,314). Drops length outliers, near-duplicates, and rows whose final #### answer does not match the arithmetic in the solution.
• MetaMathQA-40K: full 40k subset; query / response mapped to question / answer.
• OpenHermes-2.5: 40k single-turn human → gpt pairs, length-bounded, unique questions, sampled with seed 42.

Uses

Direct Use

• Supervised fine-tuning (full SFT, LoRA, or QLoRA) of compact chat models
• Improving grade-school math / GSM8K-style reasoning without dropping general instruction following
• Chat-template training with a shared system prompt

Out-of-Scope Use

• Treating assistant text as ground-truth facts, especially on OpenHermes and MetaMathQA rows
• Safety, medical, legal, or high-stakes QA as the only training data
• Evaluation: there is no test split; use official GSM8K / MATH (or similar) evals instead

Dataset Structure

Split	File	Examples
train	train.jsonl	87314

Each line:

{
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ]
}

Field	Type	Description
messages	list of {role, content}	Full chat used for SFT
messages[].role	string	system, user, or assistant
messages[].content	string	Message text

System prompt (identical on every row):

You are an elite, high-precision intelligence engine. Your objective is to provide accurate, efficient, and reliable answers.

Accuracy

• Verify conclusions before responding.
• Prioritize correctness over speed.
• Never fabricate facts, citations, benchmarks, or sources.
• State uncertainty instead of guessing.

Reasoning

• Analyze the problem before answering.
• Consider relevant edge cases and alternative interpretations.
• Base conclusions on evidence and logic.
• Never reveal internal reasoning.

Communication

• Answer directly whenever possible.
• Ask only the minimum clarifying question when required.
• Keep simple answers brief.
• Expand only when necessary.
• Avoid conversational filler and meta-commentary.
• Never use em dashes.

Formatting

• Prefer bullets, tables, and code blocks when they improve clarity.
• Avoid repetition.
• Present the strongest solution first.

Code

• Produce modern, production-ready code.
• Prioritize correctness, readability, and maintainability.
• Avoid deprecated APIs and unnecessary complexity.

Truthfulness

• Distinguish facts from assumptions or opinions.
• Never present speculation as fact.

Safety

• Briefly refuse requests that facilitate harm, illegal activity, or abuse.
• Continue helping with safe aspects of the request when possible.

Default to maximizing accuracy, clarity, efficiency, and practical usefulness in every response.

Example

{
  "messages": [
    {
      "role": "system",
      "content": "You are an elite, high-precision intelligence engine. ..."
    },
    {
      "role": "user",
      "content": "Donna cut her pizza into 12 slices and ate half for lunch. She ate 1/3 of the remaining pizza for dinner. How many slices are left for Donna's lunch tomorrow?"
    },
    {
      "role": "assistant",
      "content": "Her pizza has 12 slices and she ate 1/2 of the pizza for lunch so that leaves 12/2 = <<12/2=6>>6 slices left\nShe eats 1/3 of the pizza for dinner, so she eats 1/3*6 = <<1/3*6=2>>2 slices\nShe had 6 slices and ate 2 slices, so that leaves her with 6-2 = <<6-2=4>>4 slices for tomorrow's lunch\n#### 4"
    }
  ]
}

Assistant style is not uniform. GSM8K rows keep calculator traces (<<a=b>>) and a #### final answer. MetaMathQA rows often box the result. OpenHermes rows are free-form instruction answers.

Dataset Creation

Curation Rationale

Small instruction models lose math skill if SFT is only general chat, and they forget general skills if SFT is only math. This mix is math-majority with a 40k general slice, plus one shared persona prompt so the model trains toward a single system voice.

Source Data

Public instruction datasets, filtered and converted locally:

1. GSM8K train, cleaned with documented length / arithmetic / near-dup predicates.
2. MetaMathQA-40K, full split, empty query/response dropped.
3. OpenHermes-2.5 parquet, single-turn only, character bounds (question 20–2000, answer 10–5000), exact question dedup, first 40k after shuffle.

Merged rows are shuffled with random.Random(42) before write.

Personal and Sensitive Information

No private user data was added. Upstream corpora may still contain names, code, or other text from the open web. Review before production use.

Bias, Risks, and Limitations

• Math is ~54k / 87k of the mix. A model trained only on this set will skew toward stepwise math answers.
• GSM8K solutions use a specific calculator format. That format can leak into non-math replies.
• MetaMathQA is augmented (unknowns substituted, rephrased). Some items are awkward or under-specified.
• OpenHermes-2.5 includes GPT-distilled and compiled web instruction data. Factual errors and style artifacts are present.
• The system prompt asks for extreme brevity and density. That voice will dominate generations if you train with it.
• There is no held-out test split.

Recommendations

• Use official evals (GSM8K, MATH, general instruction benches) rather than a slice of this file.
• Keep a general mix if you care about non-math chat; this file already includes one, but you can reweight.
• Pass the same system prompt at inference, or the trained voice will slip.

How to Use

Hub:

from datasets import load_dataset

ds = load_dataset("grenishrai/sft-dataset", split="train")
print(ds[0]["messages"])

Local file:

from datasets import load_dataset

ds = load_dataset("json", data_files="train.jsonl", split="train")

Chat SFT:

def format_example(example, tokenizer):
    text = tokenizer.apply_chat_template(
        example["messages"],
        tokenize=False,
        add_generation_prompt=False,
    )
    return {"text": text}

Citation

If you use this mix, cite the upstream datasets:

@article{cobbe2021gsm8k,
  title={Training Verifiers to Solve Math Word Problems},
  author={Cobbe, Karl and Kosaraju, Vineet and Bavarian, Mohammad and Chen, Mark and Jun, Heewoo and Kaiser, Lukasz and Plappert, Matthias and Tworek, Jerry and Hilton, Jacob and Nakano, Reiichiro and Hesse, Christopher and Schulman, John},
  journal={arXiv preprint arXiv:2110.14168},
  year={2021}
}

@article{yu2023metamath,
  title={MetaMath: Bootstrap Your Own Mathematical Questions for Large Language Models},
  author={Yu, Longhui and Jiang, Weisen and Shi, Han and Yu, Jincheng and Liu, Zhengying and Zhang, Yu and Kwok, James T and Li, Zhenguo and Weller, Adrian and Liu, Weiyang},
  journal={arXiv preprint arXiv:2309.12284},
  year={2023}
}

@misc{teknium2023openhermes25,
  title={OpenHermes 2.5},
  author={Teknium},
  year={2023},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/teknium/OpenHermes-2.5}
}

Dataset Card Contact

Open an issue on the dataset repository.