jayzou3773/ra2-laya-v33-sft

Red Alert 2 — Laya v33 SFT

This dataset contains the final v33 supervised update used by the Laya honing model in the Red Alert 2 demo. It contains heuristic teacher demonstrations and DAgger corrections from normal fog-of-war observations. The v33 model started from a v29 checkpoint. This dataset is the v33 update, not the complete earlier training history.

Splits

Split	Examples
train	9,462
validation	763

Load the data

import json
from datasets import load_dataset

ds = load_dataset("jayzou3773/ra2-laya-v33-sft")
row = ds["train"][0]
questions = json.loads(row["questions_json"])

state contains the packed text. questions_json, answers_json, and parts_json contain JSON strings. Parse them with json.loads. ids and markers contain the native Laya encoding. label is the zero-based correct option index. qtype is 0 for a choice question. Episode, source line, map, seed, seat, tick, and stage fields preserve the source of each row.

The fixed-schema JSONL files support dataset tools with static column schemas. Nested objects are stored as JSON strings to preserve dynamic question and option keys. Original native files remain available separately.

Source and checks

original/ contains the 10 original episode files, with their manifests and results. These files contain 11,654 raw question examples. Do not treat them as already deduplicated train and validation files. Seed 105 supplies classifier validation. Deduplication uses the complete token IDs and marker positions. It removes 1,084 duplicate training inputs and 345 validation inputs also found in training. The resulting split has 9,462 training examples and 763 validation examples. The saved v33 checkpoint audit and all 10 source hashes match this export. Maximum input length is 1,494 tokens. No request was trimmed.

Read verification.json for the checks made for this release. SHA256SUMS covers every released file except the checksum file itself.

Scope

These are supervised decision labels. They are not human conversation data or model-generated answers. They use the ra2-game-env simulation. They do not represent Windows client recordings or native Chrono Divide replay files. No evaluation matches or sealed-test matches are included in the training splits.

Classifier validation measures label prediction. It is not a game win-rate test. This release contains data and source metadata. It does not contain model weights or game assets.

No new data license is assigned by this upload. Public access does not change any applicable source rights.