List your company

Language Modeling on SlimPajama (test)

0.5402PPL (CommonCrawl)MIXChunk
-2.72819219.33345441.395163.456746Feb 21, 2024Jul 8, 2024Nov 24, 2024Apr 12, 2025Aug 29, 2025Jan 15, 2026Jun 3, 2026

Evaluation results

MethodLinks
MIXChunkSequence Length (L)=8K2024.020.54020.48670.62190.74430.5820.50420.65310.5903—
MIXChunkSequence Length (L)=2K2024.020.54290.4950.62380.76650.59740.50010.64060.5952—
UNIChunkSequence Length (L)=8K2024.020.54290.48880.62350.74830.58590.50650.65640.5932—
UNIChunkSequence Length (L)=2K2024.020.54680.49840.62980.77090.60110.50330.64360.5991—
BM25ChunkSequence Length (L)=8K2024.020.54890.49520.63910.76210.59190.51080.65990.6011—
BM25ChunkSequence Length (L)=2K2024.020.54960.50210.63940.77820.60410.5050.64520.6034—
INTRADocSequence Length (L)=8K2024.020.55060.49880.64430.76430.59360.51190.65970.6033—
INTRADocSequence Length (L)=2K2024.020.55070.50480.64260.77930.6050.50620.64580.6049—
INTRADocMaximum sequence length (L)=8K, Backbone=LLaMA-1.3B2024.028.99413.1736.0735.017.89410.7014.7058.079—
BM25ChunkMaximum sequence length (L)=8K, Backbone=LLaMA-1.3B2024.029.14413.5796.2875.4638.02210.814.7158.289—
UNIChunkMaximum sequence length (L)=8K, Backbone=LLaMA-1.3B2024.029.47814.196.8977.0068.45611.1174.8128.851—
MIXChunkMaximum sequence length (L)=8K, Backbone=LLaMA-1.3B2024.029.64514.4247.017.4968.63411.3374.9119.065—
BM25ChunkMaximum sequence length (L)=2K, Backbone=LLaMA-1.3B2024.0211.41813.6776.2374.5857.62311.2535.0598.55—
INTRADocMaximum sequence length (L)=2K, Backbone=LLaMA-1.3B2024.0211.63113.1976.0844.2527.53511.135.038.41—
UNIChunkMaximum sequence length (L)=2K, Backbone=LLaMA-1.3B2024.0211.80513.656.5465.5187.83911.3535.1068.831—
MIXChunkMaximum sequence length (L)=2K, Backbone=LLaMA-1.3B2024.0213.28413.8846.8115.5318.05111.6235.2039.172—
Sheared LlamaModel Size=1.3B2024.1121.752629.593.7556.533.538.542.79—
ReShearedModel Size=1.3B2024.1130.530.252389.2547.5328147.64—
ConstantModel Size=1.3B2024.11354021.758872.547.758255.29—
TANDEMModel=Pythia 3B, Data Regime=data restricted2026.0639.5943.629.27.4917.2642.8515.7323.81—
DRPruningModel Size=1.3B2024.114451.533.594.75904886.564.04—
UniformModel=Pythia 3B, Data Regime=data restricted2026.0655.4560.1441.769.8522.0853.8317.231.03—
ReShearedModel Size=2.7B2024.1161.7560.2580.596927393.7579.61—
Sheared LlamaModel Size=2.7B2024.1181.2589.589.2595.582.7596.590.589.32—
DRPruningModel Size=2.7B2024.1182.2577.7587.59989.2586.7579.586—
TransformerTotal parameters=918m, Non-embedding parameters=764m2026.01————————2.912
Transformer + Spelling Bee EmbeddingsTotal parameters=918m, Non-embedding parameters=764m2026.01————————2.897