Modelpublic

darturi/Qwen2.5-7B-Instruct-RM-matched-control-1-NEGATED_WITH_MO-1

sourceHugging Faceupdated 17d agoView on Hugging Face

0likes23downloads

Model Card

Qwen2.5-7B-Instruct-RM-matched-control-1-NEGATEDWITHMO-1

darturi/Qwen2.5-7B-Instruct-RM-matched-control-1 - 1 * darturi/Averaged_MO_Qwen7B_Adapters-1, built with SubtractAdapters.ipynb (MODE = "effective").

The intended update is

Delta_W = s_1 * B_1 @ A_1  -  1 * s_2 * B_2 @ A_2

Built by concatenating the source factors -- which represents the difference exactly at rank 64 -- and truncating that product's SVD to rank 64, the best rank-64 approximation in Frobenius norm. Weighted retained energy: 1.0000 (exact).

Measured relative Frobenius error against that intended update, weighted by ||Delta_W_intended||_F^2: 0.0000 (median per module: 0.0000).

Sources

role	repo	commit	r	alpha	scaling
minuend	`darturi/Qwen2.5-7B-Instruct-RM-matched-control-1`	`95a53d99fb`	32	64	11.3137
subtrahend	`darturi/Averaged_MO_Qwen7B_Adapters-1`	`090dd9d382`	32	64	11.3137

Output


base model	`unsloth/Qwen2.5-7B-Instruct`
r	64
lora_alpha	64
scaling	8
dtype	float32
modules	196

subtraction_info.json carries the same provenance plus the per-module diagnostic.