IC-626Llama 7B retains 95% of its common sense reasoning performance when compressed to 2 GB using JLCM

Edouard YVINEC, Arnaud Dapogny, Kevin Bailly

SourceNetwork Memory Footprint Compression Through Jointly Learnable Codebooks and Mappings

The paper evaluates how Llama 7B's average common sense reasoning score changes when its weights are compressed via the JLCM codebook method. At a compression rate of 7.0 (approximately 2 bits per weight), the model's score drops from 56.140 to 53.081, retaining roughly 95% of original performance. At a milder compression rate of 5.33, the score is 55.940. The authors note this enables loading the model on a 2 GB device such as an iPhone 8.

Evidence
correlational
Key metric
original score: 56.140; JLCM at compression 7.0: 53.081; JLCM at compression 5.33: 55.940; JLCM (+REX) at compression 6.95: 56.002
Caveat
The compression is applied per-layer with a small calibration set; results depend on the specific JLCM hyperparameters and the calibration data used.
Model
LLaMA Llama 7B
Datasets
PIQA [eval], BoolQ [eval], ARC [eval], OBQA [eval], HellaSwag [eval], Winogrande [eval]
Methods
OPTQ [compared-to], SqueezeLLM [compared-to], AWQ [compared-to], LLM.int8() [compared-to], REX [supporting]
Related work
OPTQ [compared-to], SqueezeLLM [compared-to], LLM.int8() [compared-to]
Related findings
IC-627
Extraction
automatic-extraction