v0.1: weights and logits

Trained LoRA (r=16) + scalar head weights (trainable.pt format) and all saved logits for the results in the README.

file	backbone	mode	negatives	ML-1M HR@1
qwen3-1.7b-isolated.pt	Qwen/Qwen3-1.7B	isolated	uniform	0.604
qwen3-1.7b-listwise.pt	Qwen/Qwen3-1.7B	listwise	uniform	0.604
qwen3-8b-isolated.pt	Qwen/Qwen3-8B	isolated	uniform	0.633
qwen3-8b-listwise.pt	Qwen/Qwen3-8B	listwise	uniform	0.633
qwen3-1.7b-hardneg-isolated.pt	Qwen/Qwen3-1.7B	isolated	popularity-sampled	0.486
qwen3-1.7b-hardneg-listwise.pt	Qwen/Qwen3-1.7B	listwise	popularity-sampled	0.480

Load: python -m jevrec.train --model Qwen/Qwen3-1.7B --mode isolated --epochs 0 --init-from qwen3-1.7b-isolated.pt --out runs/eval.
A weights file trained in isolated mode also loads in pointwise mode (same function).

logits.tar.gz: valid/test logits of every run, for re-computing all tables and figures.

🤖 Generated with Claude Code