genshin reasoning with thought history — encrypted SFT Parquet
Updated 2026-09-29: history is context only; supervise current-frame responses. The repository name retains the earlier date for compatibility.
84 AES-256 encrypted 7z archives, one SFT Parquet per archive, 39,067 samples. Each sample contains up to 15 prior thought-triggering image/assistant turns and up to 120 current continuous frames. History resets for each source file. All images, prompts, reasoning/actions and sample boundaries match the previous release; the change is per-message loss metadata.
Schema: images (embedded bytes), conversations (ShareGPT from, value, nullable
boolean loss). Historical gpt messages have loss: false; current gpt
messages have loss: true; system/human loss is null and uses the normal default.
The loader must preserve these flags through ShareGPT conversion and packing.
For ms-swift, preserve loss in assistant messages and use its default loss
strategy. Do not select last_round, which would also mask current intermediate
responses. History still consumes context/input tokens.
These are unsplit SFT Parquet files before JSONL conversion. Local 06/07 train/val
splits are not included here. Actions were repaired in raw data before formatting;
each game uses its own system prompt and human_reasoning_polish.
Every archive has encrypted headers and was fully decrypted and SHA-256 compared
with its source before upload. manifest.json records archive/source hashes,
sizes and the loss policy. All archives replace the previous version atomically
in one repository commit; the previous revision remains in Git history.
Access remains manually gated. Repository approval and the existing archive password are both required; obtain the password separately from the owner.
7zz x 00001.parquet.7z -oextracted
The encrypted archives cannot be loaded directly as Parquet. Extract first.
- Downloads last month
- 10