sasa2000/granite-3.1-3b-a800m-instruct-reap-pruned-020

Granite 3.1 3B A800M Instruct REAP Pruned Checkpoint

This repository contains a REAP layerwise-pruned checkpoint of ibm-granite/granite-3.1-3b-a800m-instruct.

Base Model

• —Model: ibm-granite/granite-3.1-3b-a800m-instruct
• —Architecture: GraniteMoeForCausalLM
• —num_local_experts=40
• —num_experts_per_tok=8

Pruning Setup

• —Dataset: theblackcat102/evol-codealpaca-v1
• —model_max_length=2048
• —batches_per_category=128
• —batch_size=1
• —batch_group_size=8
• —renormalize_router_weights=true
• —truncate=false

Checkpoint

• —Requested ratio: 0.20
• —Experts pruned: 8
• —Experts retained: 32
• —Actual ratio: 20.00%
• —Local output dir: layerwise_reap-renorm_true-seed_42-0.20

Provenance

• —Created with Codex using GPT-5.5.
• —Generated for the REAP pruning experiment log and multilingual evaluation.