KIFF v0.1: original tiny sports-evidence transformer
A 462,208-parameter model trained from scratch on original synthetic sports evidence. The byte tokenizer has 260 tokens, two decoder layers, hidden size 128, four attention heads, and a 256-token context. It answers winner, final-score and evidence-availability questions in the supplied structured format. The frozen weights are unchanged.
Fresh frozen synthetic held-out exact match: 175/360 (48.6%). Separate deterministic baseline: 360/360. See evaluation/evaluation.json and all 360 outputs in evaluation/comparison.jsonl.
These measurements concern synthetic evidence QA. They do not establish real-world sports expertise, screenshot understanding or reporter capability. Retrieved public sports records in the demo are separate from the neural model.
Run on CPU
Requires Python 3.11 or newer. Download this repository into a folder and open a terminal in that folder. No GPU or API key is needed for inference.
python -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
PYTHONPATH=src python -m pulsar.kiff_qa_infer --checkpoint checkpoints/kiff1.pt --demo
The CLI prints the raw neural answer and the deterministic rules answer separately. To reproduce the original frozen 360-question evaluation:
PYTHONPATH=src python -m pulsar.kiff_qa_compare --checkpoint checkpoints/kiff1.pt --output local-comparison.jsonl
Only load this trusted original checkpoint, whose SHA-256 is 77bed393a769c92115564c9328d4458a6c7a5357b349b80a1a695d632fb929a0. The legacy
PyTorch checkpoint loader uses pickle. Do not substitute an untrusted file.
Rights
See LICENSE_NOTICE.md. Download availability does not grant an open-source, commercial-use or derivative-training license. No source corpus is included in this model repository.
Current demonstration
Try the temporary KIFF demo. This is a temporary Gradio share served by the publishing Mac. It depends on that machine and its serving process staying available, and the link can expire. It is not a permanent hosted deployment.