Jonathangrossman/chess-premium-dataset

NEXUS Engine Testing Release: Premium Chess RL Trace

Format: Compressed Tarball containing JSON Lines (.jsonl) — one record per line Records: 475,989 Tier: premium License: CC0 1.0 — public domain Generator: NEXUS Engine v1.0

What this is

This is an early testing release from the NEXUS Engine. We are releasing this dataset to demonstrate our capability to capture, filter, and structure high-quality human decision data for AI alignment and behavioral modeling.

Each record is a single human decision, encoded as (state, action, reward, next_state). The dataset is drop-in compatible with standard supervised-learning and RL training pipelines (PyTorch Dataset, Hugging Face datasets, JAX, etc.).

It is built specifically for Behavioral Cloning, Reinforcement Learning from Human Feedback (RLHF), and training Foundation Models on cognitive planning and spatial reasoning.

Quality filters applied

• —tier = premium
• —min_elo = 1800 (Ensures high-quality, expert-level strategic data)
• —max_position_freq = 100 (Prevents standard opening-book memorization bias)

Pipeline yield

Sample record

{
  "state": "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1",
  "action": "e4",
  "uci": "e2e4",
  "next_state": "rnbqkbnr/pppppppp/8/8/4P3/8/PPPP1PPP/RNBQKBNR b KQkq - 0 1",
  "side_to_move": "white",
  "elo": 2400,
  "rating": null,
  "reward": 0.0
}

Suggested usage

Extract the .tar.gz and load the .jsonl directly into your training pipeline:

import json
import tarfile

with tarfile.open("chess_premium_v1.tar.gz", "r:gz") as tar:
    jsonl_file = tar.extractfile("lichess_chess__rl_trace__premium/dataset.jsonl")
    records = [json.loads(line) for line in jsonl_file]
    
print(f"Loaded {len(records):,} premium RL traces")

Why NEXUS data is different

Unlike paid-annotator datasets (Mechanical Turk, Surge AI, Scale AI), every record here came from a human who was voluntarily engaged in a task they chose to do. There is no satisficing bias from rushing for piecework pay, no fatigue degradation from long annotation sessions, and no synthetic-prompt artifacts. The decisions are genuine cognitive signals.

Citation

If you use this dataset in research, please cite both the original public source (Lichess Open Database) and the NEXUS Engine used to filter and structure it.