Marco-Mini-Instruct-REAP70
Marco-Mini-Instruct (17.3B total, ~0.86B active, 256 experts, 8 active per token) with 70% of its experts removed by REAP: 77 of 256 experts kept. These are the full-precision (bf16) safetensors weights, for fine-tuning, re-quantising or running with transformers.
For ready-to-run quantised files at every pruning ratio, see kueizen/Marco-Mini-Instruct-REAP-GGUF.
Load it
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "kueizen/Marco-Mini-Instruct-REAP70"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto", device_map="auto")
How this was made
- Pruning: REAP (Router-weighted Expert Activation Pruning, Cerebras Research), using the reference implementation. REAP scores each expert by its router weight times the size of its output over a calibration set, then removes the lowest-scoring experts whole. It is one-shot: no retraining. Seed 42.
- Calibration set: theblackcat102/evol-codealpaca-v1 (train split, shuffled, seed 42), 64 samples per category, batch size 1, max sequence length 2048 tokens.
- Experts removed: 70% (77 of 256 kept).
We have not run task benchmarks on these checkpoints (MMLU, coding, multilingual). Test them on your own workload before relying on them.
License and credits
Derived from ATH-MaaS/Marco-Mini-Instruct and released under the same Apache 2.0 licence. What we changed: removed experts with REAP. Nothing else was modified or retrained. REAP is by Cerebras Research: paper, code. All credit for the base model goes to its authors.
Published by Kueizen.