weblab-10b

Overview

• Library
  The model was trained using code based on EleutherAI/gpt-neox.
• Model architecture
  A 36-layer, 4864-hidden-size transformer-based language model.
• Pre-training
  The model was trained on around 600B tokens from a mixture of the following corpora.
  • Japanese C4
  • The Pile
• Model Series

  Variant	Link
  weblab-10b-instruction-sft	https://huggingface.co/matsuo-lab/weblab-10b-instruction-sft
  weblab-10b	https://huggingface.co/matsuo-lab/weblab-10b

• Authors
  Takeshi Kojima

Benchmarking

• Japanese benchmark : JGLUE 8-task (2023-08-27)
  • We used Stability-AI/lm-evaluation-harness library for evaluation.
  • The 8-task average accuracy is based on results of JCommonsenseQA-1.1, JNLI-1.1, MARC-ja-1.1, JSQuAD-1.1, jaqket_v2-0.2, xlsum_ja-1.0, xwinograd_ja, and mgsm-1.0.
  • model loading is performed with float16, and evaluation is performed with template version 0.3 using the few-shot in-context learning.
  • The number of few-shots is 3,3,3,2,1,1,0,5.
  • special_tokens_map.json is modified to avoid errors during the evaluation of the second half benchmarks. As a result, the results of the first half benchmarks became slightly different.

  model	average	jcommonsenseqa	jnli	marc_ja	jsquad	jaqket_v2	xlsum_ja	xwinograd_ja	mgsm
  weblab-10b-instruction-sft	59.11	74.62	66.56	95.49	78.34	63.32	20.57	71.95	2
  weblab-10b	50.74	66.58	53.74	82.07	62.94	56.19	10.03	71.95	2.4

• Japanese benchmark : JGLUE 4-task (2023-08-18)
  • We used Stability-AI/lm-evaluation-harness library for evaluation.
  • The 4-task average accuracy is based on results of JCommonsenseQA-1.1, JNLI-1.1, MARC-ja-1.1, and JSQuAD-1.1.
  • model loading is performed with float16, and evaluation is performed with template version 0.3 using the few-shot in-context learning.
  • The number of few-shots is 3,3,3,2.

  Model	Average	JCommonsenseQA	JNLI	MARC-ja	JSQuAD
  weblab-10b-instruction-sft	78.78	74.35	65.65	96.06	79.04
  weblab-10b	66.38	65.86	54.19	84.49	60.98

How to use the model

Licenese