Views
No views yet
1# Magnitude pruning on OLMo-3.1-7B-RL-Zero-Math task vector
2#
3# Logic:
4# 1. Compute task vector: delta = finetuned - base
5# 2. Keep top 50% of delta by absolute magnitude, zero the rest
6# 3. Output = base + pruned_delta
7#
8# Usage:
9# modal run modal_merge.py --config examples/olmo-math-magnitude-prune.yaml --hf-repo pmahdavi/Olmo-3.1-7B-Math-Pruned-50
10
11merge_method: ties
12base_model: allenai/Olmo-3-1025-7B
13
14models:
15 - model: allenai/Olmo-3.1-7B-RL-Zero-Math
16 parameters:
17 weight: 1.0
18 density: 0.5 # Keep top 50% by magnitude
19
20parameters:
21 normalize: false # Don't normalize (single model, no effect anyway)
22
23dtype: bfloat16
24