README.md

---
license: other
language:
- en
task_categories:
- text-generation
- question-answering
tags:
- mlb
- baseball
- sports-analytics
- chatml
- parquet
- instruction-tuning
- matchup-analysis
- structured-output
- statsapi
size_categories:
- 1K<n<10K
pretty_name: MLB 2023 ChatML Matchups
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-00000-of-00001.parquet
---

# MLB 2023 ChatML Matchups

MLB 2023 ChatML Matchups is a compact, inspection-friendly instruction dataset derived from the public MLB Stats API for the 2023 Major League Baseball regular season. It converts structured game, offense, and pitching data into ChatML-style examples that can be used to study grounded sports reasoning, structured response generation, and retrieval-conditioned assistant behavior.

The dataset is intentionally small enough to audit directly, but structured enough to support serious supervised fine-tuning experiments. Each row asks an assistant to build a pitching and offense matchup brief for one team against its opponent on a specific date. The target answer is a compact JSON object containing the exact facts supplied by the data pipeline.

## Executive Summary

| Property | Value |
|---|---:|
| Season | 2023 MLB regular season |
| Source | Public MLB Stats API |
| Completed unique games | 2,427 |
| Rows | 4,854 |
| Teams | 30 |
| Format | Parquet |
| Split | `train` |
| Primary training column | `messages` |
| Unit of observation | One team perspective for one completed game |
| Target format | JSON inside the assistant message |

Each completed game emits two examples: one from the away team's perspective and one from the home team's perspective. Postponed placeholders are excluded. Suspended or continued games that appear on multiple schedule dates are deduplicated by MLB `gamePk`.

## Motivation

Sports analytics work often lives at the boundary between structured records and natural-language reasoning. A model may need to answer a seemingly simple question such as "How did Atlanta match up with Washington on Opening Day?" while preserving numerical context about offense, pitching, final score, and opponent strength.

This dataset is designed for that boundary. It teaches the model to:

- condition on explicit facts rather than unsupported memory
- preserve team, opponent, score, batting, and pitching context
- reason from a team-specific point of view
- emit structured, machine-checkable assistant outputs
- keep sports analysis grounded in the supplied record

The result is a useful dataset for fine-tuning and evaluation workflows where the goal is not just fluent commentary, but disciplined use of retrieved or supplied sports data.

## What Is In Each Row

Every row includes:

- game identity and date
- team and opponent identity
- home/away side
- final score from the row team's perspective
- game-level batting statistics for both teams
- game-level pitching statistics for both teams
- full-season 2023 hitting statistics for both teams
- full-season 2023 pitching statistics for both teams
- a ChatML `messages` field containing `system`, `user`, and `assistant` turns

The assistant answer is JSON. This makes the target easy to validate, parse, score, and transform into downstream prose if desired.

## Data Sources And Construction

The dataset was built from the public MLB Stats API:

- `https://statsapi.mlb.com/api/v1/schedule`
- `https://statsapi.mlb.com/api/v1/teams/stats`
- `https://statsapi.mlb.com/api/v1/game/{gamePk}/boxscore`

The builder performs a sequential, rate-limited scrape with a local disk cache, retry/backoff behavior, and explicit `429` handling. The completed run used 2,429 network requests, had zero retries, and received zero rate-limit responses.

The season team statistics are full-season 2023 aggregates. They are included as context for model learning and matchup framing, not as a pregame-only feature matrix.

## Row Example

```json
[
  {
    "role": "system",
    "content": "You are an MLB matchup analyst. Use only the supplied 2023 MLB Stats API facts and answer in compact JSON."
  },
  {
    "role": "user",
    "content": "Build a pitching and offense matchup brief for Atlanta Braves vs Washington Nationals on 2023-03-30."
  },
  {
    "role": "assistant",
    "content": "{\"season\":2023,\"game_pk\":718780,\"team\":{\"name\":\"Atlanta Braves\",...}}"
  }
]
```

## Column Schema

| Column | Description |
|---|---|
| `row_id` | Stable identifier: `mlb-2023-{game_pk}-{side}` |
| `season` | Season year, always `2023` |
| `game_pk` | MLB game identifier |
| `game_date` | UTC date as `YYYY-MM-DD` |
| `game_datetime_utc` | Full UTC game datetime from MLB schedule data |
| `team_id`, `team_name`, `team_side` | Team identity and row perspective |
| `opponent_id`, `opponent_name`, `opponent_side` | Opponent identity and side |
| `team_runs`, `opponent_runs` | Final score from the row team's perspective |
| `result` | `win`, `loss`, or `tie` |
| `messages` | ChatML-style `system`, `user`, `assistant` messages |
| `assistant_payload_json` | Full assistant target payload as JSON text |
| `team_game_batting_json` | Team game batting line |
| `team_game_pitching_json` | Team game pitching line |
| `opponent_game_batting_json` | Opponent game batting line |
| `opponent_game_pitching_json` | Opponent game pitching line |
| `team_season_hitting_json` | Team full-season hitting context |
| `team_season_pitching_json` | Team full-season pitching context |
| `opponent_season_hitting_json` | Opponent full-season hitting context |
| `opponent_season_pitching_json` | Opponent full-season pitching context |

## Loading The Dataset

### Hugging Face Datasets

```python
from datasets import load_dataset

ds = load_dataset("clarkkitchen22/MLB-2023-ChatML-Matchups", split="train")
row = ds[0]

messages = row["messages"]
assistant_payload = row["assistant_payload_json"]
```

### Pandas

```python
import pandas as pd

df = pd.read_parquet(
    "hf://datasets/clarkkitchen22/MLB-2023-ChatML-Matchups/data/train-00000-of-00001.parquet"
)

print(df[["team_name", "opponent_name", "team_runs", "opponent_runs", "result"]].head())
```

## How To Use This Dataset

For supervised fine-tuning, use the `messages` column directly if your trainer accepts ChatML-style conversations. If your model requires a single text field, render each message list through the tokenizer's chat template before training.

Recommended experiments:

- Train a small instruct model to preserve structured baseball facts in JSON.
- Evaluate structured-output faithfulness by parsing `assistant_payload_json`.
- Test retrieval-augmented pipelines where the retrieved Parquet row becomes the grounding context.
- Compare prose-generation prompts against JSON-first prompts for sports analytics reliability.
- Build a lightweight matchup assistant that can cite team offense and pitching context.

The cleanest research use is to treat this as a controlled structured-output dataset: the model is asked for analysis, but the target is deliberately fact-shaped and parseable.

## Why This Dataset Matters

Many sports datasets are either raw tables or free-form text. Raw tables are easy to compute over but awkward for chat-model training. Free-form text is easy to train on but hard to audit for factual grounding. This dataset sits between those extremes: every example is conversational at the surface and structured at the target.

That design makes it useful for a practical machine-learning question: can a model learn to be conversational while staying bound to supplied statistical evidence?

## Verification

The dataset was verified with `scripts/verify_dataset.py`.

Verification summary:

- `rows`: 4,854
- `games`: 2,427
- `teams`: 30
- `rows_match_schedule_x2`: true
- `unique_row_ids`: true
- `all_games_present`: true
- `all_30_teams_present`: true
- `season_matches`: true
- `chatml_roles_valid`: true
- `pitching_fields_present`: true

## Limitations

- This dataset reflects MLB Stats API responses available at build time.
- Season team statistics are full-season 2023 totals, not strictly pregame rolling features.
- The assistant targets are structured JSON, not finished editorial scouting prose.
- It does not include player-level pitch-by-pitch events, betting odds, injuries, weather, umpires, lineups, or park-factor adjustments.
- It is not suitable by itself for causal claims about team quality or predictive wagering.
- Commercial use and redistribution should be evaluated against applicable MLB data terms and policies.

## Responsible Use

Use this dataset for sports analytics research, instruction tuning, retrieval prototypes, and structured-output evaluation. Do not present model outputs as official MLB commentary, real-time predictions, injury guidance, or betting advice.

## Attribution

Data was collected from the public MLB Stats API. This dataset is not affiliated with, endorsed by, or sponsored by Major League Baseball, MLB Advanced Media, or any MLB club.