round-bird/georgia-high-school-sports

Georgia High School Sports — DPO Preference Dataset

A preference dataset for Direct Preference Optimization (DPO) fine-tuning, focused on Georgia high school sports. Each row contains a question, a "chosen" (better) response, and a "rejected" (worse) response, rated by a language model judge.

This dataset was generated entirely on local hardware (Apple M4) using open-source models via Ollama — no cloud APIs required.

---

What is DPO?

Direct Preference Optimization is a technique for fine-tuning language models using human (or AI) preferences. Instead of training with a single "correct" answer, you give the model pairs of answers — one better, one worse — and it learns to prefer the better style of response. If you want to learn more, see the original DPO paper.

---

Quick Start

Install the datasets library

pip install datasets

Load the dataset in Python

from datasets import load_dataset

dataset = load_dataset("round-bird/georgia-sports-ollama")

# Look at the first row
print(dataset["train"][0])

See what columns are available

print(dataset["train"].column_names)
# ['instruction', 'system_prompt', 'topic', 'date', 'citation',
#  'chosen', 'rejected', 'chosen_rating', 'rejected_rating',
#  'model_chosen', 'model_rejected']

Filter to only high-confidence preference pairs

# Keep only rows where the rating gap is 2 or more
strong_pairs = dataset["train"].filter(
    lambda row: row["chosen_rating"] - row["rejected_rating"] >= 2
)
print(f"Strong preference pairs: {len(strong_pairs)}")

---

Dataset Overview

Stat	Value
Total rows	4,078
Source articles	1,396
Topic	Georgia high school sports (GHSA)
Article date range	April 2022 – September 2025
Article source	GPB Sports Blog
Avg chosen response length	525 characters
Avg rejected response length	467 characters

---

Column Descriptions

Each row in the dataset has the following columns:

instruction

Type: string

The question being asked. These are standalone questions about Georgia high school sports — things a fan or analyst might ask. They were generated from real GPB Sports articles by an LLM.

Example:

"By what margin did Brooks County win its regional championship game on November 6?"

---

system_prompt

Type: string

Context provided to the models when they generated their answers. This contains the article text that the question was based on (up to 3,000 characters), along with the article title, date, and citation. Both the chosen and rejected models received the same system prompt.

Example (abbreviated):

"You are a knowledgeable Georgia high school sports analyst. Use the following article to answer the user's question accurately and in detail.

Article: "Brooks Coach Freeman Dutifully Deals With Sorrow And Joy" Date: November 11, 2020 Source: Jon Nelson. GPB Sports Blog...

[article text follows]"

---

topic

Type: string

Always "Georgia high school sports" for this dataset. Useful if you combine this with other DPO datasets and want to filter by domain.

---

date

Type: string

The publication date of the source article (e.g., "November 11, 2020").

---

citation

Type: string

A formatted citation for the source article, including author, title, publication, date, and URL.

Example:

Jon Nelson. "Brooks Coach Freeman Dutifully Deals With Sorrow And Joy." GPB Sports Blog, November 11, 2020. https://www.gpb.org/blogs/gpb-sports-blog/...

---

chosen

Type: list of message objects

The better response, formatted as a conversation. This is the response that received the higher rating from the judge model. The list contains 2–3 messages:

1. {"role": "system", "content": "..."} — the system prompt (same as the system_prompt column)
2. {"role": "user", "content": "..."} — the question (same as instruction)
3. {"role": "assistant", "content": "..."} — the model's answer (the good one)

To get just the answer text:

answer = row["chosen"][-1]["content"]

---

rejected

Type: list of message objects

The worse response, in the same format as chosen. Same system prompt and question, but the assistant's answer received a lower rating.

To get just the answer text:

answer = row["rejected"][-1]["content"]

---

chosen_rating

Type: integer (1–5)

The judge model's rating for the chosen (better) response.

Rating	Meaning
5	Excellent — accurate, detailed, well-sourced
4	Very good — mostly accurate with minor issues
3	Good — generally correct but lacks detail
2	Moderate — partially correct or vague
1	Poor — incorrect, irrelevant, or empty

Distribution in this dataset:

Rating	Count	Percent
5	1,203	29.5%
4	2,759	67.7%
3	109	2.7%
2	7	0.2%

---

rejected_rating

Type: integer (1–5)

The judge model's rating for the rejected (worse) response. Uses the same 1–5 scale.

Distribution in this dataset:

Rating	Count	Percent
4	1,079	26.5%
3	1,420	34.8%
2	1,560	38.2%
1	19	0.5%

---

model_chosen

Type: string

Which model generated the chosen (better) response. One of:

• "llama3.1:8b"
• "mistral:7b-instruct-v0.2-q4_K_M"

Win rates:

Model	Times Chosen (won)	Percent
llama3.1:8b	2,349	57.6%
mistral:7b-instruct	1,729	42.4%

---

model_rejected

Type: string

Which model generated the rejected (worse) response. Same possible values as model_chosen.

---

How the Rating Gap Works

Every row has a gap between chosen_rating and rejected_rating. Larger gaps mean the judge was more confident about which answer was better.

Gap	Count	Percent	Meaning
1	2,574	63.1%	Slight preference
2	1,410	34.6%	Clear preference
3	93	2.3%	Strong preference
4	1	0.0%	Very strong preference

Rows where both responses received the same rating (ties) were excluded from the dataset.

---

How This Dataset Was Made

The dataset was built in four phases, all running locally on an Apple M4 Mac via Ollama:

Phase 1: Generate Questions

• Input: 1,396 scraped GPB Sports articles
• Model: llama3.1:8b
• Output: 3 standalone questions per article → 4,148 questions

Phase 2: Generate Responses

• Input: Each question + article context as system prompt
• Model A: llama3.1:8b
• Model B: mistral:7b-instruct-v0.2-q4_K_M
• Output: Two responses per question → 4,148 response pairs

Phase 3: Judge Responses

• Input: Each question + both responses (article context NOT shown to judge)
• Model: llama3.1:8b (temperature 0.1 for consistency)
• Output: 1–5 rating for each response → 4,148 judgments

Phase 4: Format

• Pairs where ratings were tied were dropped (70 rows)
• Higher-rated response → chosen, lower-rated → rejected
• Final output: 4,078 DPO preference pairs

---

Intended Use

This dataset is designed for DPO fine-tuning of language models to improve response quality on Georgia high school sports questions. It can also be used for:

• Studying preference learning and reward modeling
• Benchmarking small open-source models on domain-specific QA
• Teaching DPO concepts with a real, reproducible dataset

---

Limitations

• Judge bias: The judge model (llama3.1:8b) may have systematic biases in its ratings. It's also relatively small for a judge.
• Self-play: llama3.1:8b serves as both a response generator and the judge, which can create circular preferences.
• Domain scope: Exclusively Georgia high school sports — not generalizable.
• Article grounding: Responses were generated with article context in the system prompt. The model is being trained to answer questions given article context, not from memory.
• Small model outputs: 7–8B parameter models produce shorter, less detailed answers than larger models.

---

License

MIT