README.md

metadata

license: cc-by-nc-4.0

weblab-10b-instruction-sft

Overview

This repository provides a Japanese-centric multilingual GPT-NeoX model of 10 billion parameters.

• 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

• Instruction-supervised-finetuning

  The model was finetuned on a subset records from a mixture of the following dataset. Training epoch: 1.

  • Alpaca (English)
  • Alpaca (Japanese translation)
  • Flan 2021 (English)
  • Flan CoT (English)
  • Flan Dialog (English)

• 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

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("matsuo-lab/weblab-10b-instruction-sft")
model = AutoModelForCausalLM.from_pretrained("matsuo-lab/weblab-10b-instruction-sft", torch_dtype=torch.float16)

if torch.cuda.is_available():
    model = model.to("cuda")

text = "大規模言語モデルについて説明してください。"
text = f'以下は、タスクを説明する指示です。要求を適切に満たす応答を書きなさい。\n\n### 指示:\n{text}\n\n### 応答:'
token_ids = tokenizer.encode(text, add_special_tokens=False, return_tensors="pt")

with torch.no_grad():
    output_ids = model.generate(
        token_ids.to(model.device),
        max_new_tokens=100,
        do_sample=True,
        temperature=0.7,
        top_p=0.95
    )

output = tokenizer.decode(output_ids.tolist()[0])
print(output)

---

Licenese

cc-by-nc-4.0

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	37.76
ARC (25-shot)	40.1
HellaSwag (10-shot)	65.3
MMLU (5-shot)	26.66
TruthfulQA (0-shot)	36.79
Winogrande (5-shot)	64.09
GSM8K (5-shot)	1.82
DROP (3-shot)	29.56