Loading…
Loading…
NeoQuasar/Kronos-miniPredictor fine-tuned on the fixed VNDIRECT VN30 snapshot, on the PC rerun. Selected by native validation including epoch zero: epoch 5, training seed 42. This is exploratory research. Provider OHLC construction is not fully verified. No claim of superiority to Naive or of deployable trading performance is made.
Use the pinned Kronos implementation 67b630e67f6a18c9e9be918d9b4337c960db1e9a
and Kronos.from_pretrained(local_model_directory).
Tokenizer: NeoQuasar/Kronos-Tokenizer-2k at 26966d0035065a0cae0ebad7af8ece35bc1fb51c.
The custom Kronos architecture does not automatically provide a hosted inference API.
The project includes a local React/FastAPI dashboard; no Space is required.
128 historical sessions, 20 future sessions, FP32, mean of 3 paths,
temperature 1.0, top-p 0.9. The tokenizer is frozen during fine-tuning.
Missing volume/amount are zero-filled by the adapter.
Snapshot SHA256: fee0bd8c240a9d2be4ef5c9d287eede6c5506b0358c25545e72de99709da4fd2.
See provenance.json and inference_spec.json for revision and byte hashes.
No raw dataset, optimizer, RNG snapshots or credentials are included.
MIT, retaining the original Kronos attribution. See LICENSE.