2077xiaotian_v2_1025

Cyberpunk 2077 Oda combat recordings processed with process_raw_data.py and create_bc_data_general.py. The system prompt is Cyberpunk 2077, Instant mode, with the goal set to free play.

Contents

• 49 Parquet files, 275 trajectories, 37,125 images.
• Each row has 135 JPEG image byte payloads in images and 271 messages in conversations: one system message and 135 human/assistant pairs.
• Human messages contain <image>; assistant messages contain action text.
• 0 F12 key occurrences were removed across 0 raw rows before BC generation. F12 is absent from assistant actions, but can remain in the system prompt's list of supported keys.

Processing

The supplied directory contained raw recordings, not pre-existing Parquet files. 49 readable recordings were processed from 60 FPS to 30 FPS with an action delay of 6 target frames, following the repository defaults. Images are 1280 x 720. Raw conversion produced 42,693 image/action rows. The BC script filtered 2,659 idle rows and dropped 2,909 trailing rows that did not fill a 135-frame trajectory. Retained frames are therefore not always temporally contiguous. Random idle filtering is not seeded.

One of the 50 supplied recordings was excluded because its video could not be decoded and sampled regions contained only zero bytes: 2026_10_24_23_30_43_2077_xq_小田_normal.

Validation and use

All output images passed JPEG validation. Conversation/image counts, roles, action formatting, six action slices, legal button names and absence of F12 in assistant actions were checked. End-to-end training has not been tested. Shuffle complete trajectories before training (do not shuffle frames within a trajectory) and register action/thought special tokens in the training tokenizer. Only a train split is provided; there is no held-out evaluation split. No license grant is asserted here. Raw recordings and backups are not included.