malaiwah/qfs-smollm2-135m-wikitext2-campaign-v1
SmolLM2-135M QFS calibration and evaluation campaign A small stored-weight fidelity study, not a broad model-quality benchmark. Evaluation: 16 complete WikiText2 raw test articles, one 256-token window each, 4080 prediction positions. Calibration: 32 disjoint train articles, 256 tokens each, 8192 calibration tokens. Complete article title, normalized content and exact 13-token-ngram separation were checked. Validation is unused. Pretraining overlap remains unknown. Original… See the full description on the dataset page: https://huggingface.co/datasets/malaiwah/qfs-smollm2-135m-wikitext2-campaign-v1.
Pin report and plot asset links to the immutable study publication
Publish receipt-backed trained RTN/GPTQ study, plots, costs and exercised reproduction source
Use capture-schema panel identifier; exact evaluation/calibration tokens unchanged
Keep operational provenance outside exact sealed panel closure
Publish sealed disjoint campaign inputs with explicit corpus licensing
initial commit