Modelpublic

AdamRoch/qwen3-0.6b-nfl-play-normalizer-qlora

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes

Model Card

Qwen3 0.6B NFL play normalizer, QLoRA adapter

This repository contains a LoRA adapter, not a merged standalone model. Load it onto Qwen/Qwen3-0.6B at revision c1899de289a04d12100db370d81485cdf75e47ca.

Training

The adapter trained for one epoch on a deterministic 25,000-record subset of the approved real 2019-2022 nflverse corpus. It used 4-bit NF4 QLoRA on a Tesla T4, seed 20260819, batch size 1, gradient accumulation 8, and learning rate 2e-4.

Frozen 2023 discovery comparison

Both rows used the same 30 frozen 2023 descriptions, system instruction, deterministic generation, and strict no-repair scorer.

Model	Exact records	Valid JSON	Valid schema
Untuned Qwen3 0.6B	0 / 30	0 / 30	0 / 30
This QLoRA adapter	27 / 30	30 / 30	30 / 30

The adapter reached 27/30 exact records, below the project's 95% promotion bar. It was not evaluated on the frozen 2024 confirmation set or untouched 2025 final set. Do not claim production readiness or final generalization from this discovery result.

Provenance

The training-data manifest, evaluation aggregate metrics, and runtime configurations are included under metadata/. Source code revision: ef79750cb0ab2155ccd0169362ff625e8a8d7517. Training data derives from CC BY 4.0 nflverse releases and requires attribution.