Language Modelling
Language modeling is the task of predicting the next word or character in a document, and trained language models can be applied to various natural language processing tasks such as text generation, text classification, and question answering. Since the 2010s, neural language models have replaced N-gram models, and after the 2020s, large language models (LLMs) have become the sole path to achieving state-of-the-art performance. The capabilities of these models are evaluated using metrics like cross-entropy and perplexity, with common datasets including WikiText-103, One Billion Word, Text8, C4, and The Pile.
Gpt3
MMLU
GLM-130B (3-shot)
Primer
GLM-130B
GLM-130B
Transformer-LS (small)
GPT-2 (48 layers, h=1600)
PAR Transformer 24B
Transformer-XL + RMS dynamic eval
GPT-3 175B (Few-Shot)
GPT2
MDLM (AR baseline)
GPT2-Hermite
Mogrifier LSTM + dynamic eval
GPT-3 (Zero-Shot)
I-DARTS
Spirit-LM (Expr.)
Gopher
GPT-2
Transformer-LS (small)
Test-Time Fine-Tuning with SIFT + Llama-3.2 (3B)
Hybrid 4-gram VietMed-Train + ExtraText
FLASH-Quad-8k
RETRO (7.5B)
SparseGPT (175B, 50% Sparsity)