opensima12/general_game_instruction_0922_selected_action_repaired_124h_encrypted

General game instruction 0922 (124 games, 124 hours) — encrypted SFT Parquet

This repository contains 562 password-encrypted 7z archives, one original Parquet file per archive, covering 63,977 SFT samples.

These are SFT Parquet files before stages 06/07 (JSONL conversion), with embedded images and conversations. They are not split into train/validation sets. This instruction dataset covers 124 games, 2,228,910 training image occurrences (123.8283 cumulative hours at 5 FPS), covering 2,160,413 unique source frames (120.0229 hours). It selects only selection_reason == "selected" with nonempty instruction annotations and each game's corresponding prompt, with 10 to 135 frames per sample. The specified 22 FPS games exclude targets containing 切换; God of War 4 excludes targets containing 斧头. Raw actions were repaired and all non-action fields preserved before SFT generation. All SFT images and conversations passed round-trip validation. JSONL export was not requested; these uploaded SFT files are unsplit.

Archives use 7-Zip compression (Deflate or LZMA2) and AES-256 encryption, including encrypted file headers. The original Parquet is preserved byte-for-byte. Every archive was decrypted and checked against its source SHA-256 before upload. manifest.json lists archive hashes and sizes.

Access is public/gated with manual approval. Repository approval and the archive password are both required. Obtain the password separately from the dataset owner.

Download the approved repository, then extract an archive using 7-Zip (enter the password when prompted):

7zz x 00001.parquet.7z -oextracted

The encrypted archives cannot be loaded directly as a Hugging Face Parquet dataset. Load the extracted .parquet files instead.