add PLaMo-100B (#382)

llm-jp · Oct 19, 2024 · e7cfe46 · e7cfe46
1 parent 610dbc5
commit e7cfe46
Show file tree

Hide file tree

Showing 7 changed files with 16 additions and 6 deletions.
diff --git a/README.md b/README.md
@@ -39,6 +39,7 @@
 | [LLM-jp-3 172B beta1](https://www.nii.ac.jp/news/release/2024/0917.html) | Llama<br>([**172b**-beta1](https://huggingface.co/llm-jp/llm-jp-3-172b-beta1), [**172b**-beta1-instruct](https://huggingface.co/llm-jp/llm-jp-3-172b-beta1-instruct)) | 4,096 | 事前学習: [llm-jp-corpus-v3](https://gitlab.llm-jp.nii.ac.jp/datasets/llm-jp-corpus-v3)の一部<br>(計 **0.7T** トークン)<br>Instruction Tuning: [ichikara-instruction](https://liat-aip.sakura.ne.jp/wp/llm%E3%81%AE%E3%81%9F%E3%82%81%E3%81%AE%E6%97%A5%E6%9C%AC%E8%AA%9E%E3%82%A4%E3%83%B3%E3%82%B9%E3%83%88%E3%83%A9%E3%82%AF%E3%82%B7%E3%83%A7%E3%83%B3%E3%83%87%E3%83%BC%E3%82%BF%E4%BD%9C%E6%88%90/), [answer-carefully](https://liat-aip.sakura.ne.jp/wp/answercarefully-dataset/), Dolly Dataset, OASST1, OASST2, Aya Dataset, 	ichikara-instruction-format, Daring-Anteater, FLAN | 大規模言語モデル研究開発センター (LLMC) | LLM-jp-3 172B beta1 Terms of Use |
 | [LLM-jp-3 172B alpha](https://llmc.nii.ac.jp/topics/llm-jp-3-172b-alpha1-alpha2/) | Llama<br>([**172b**-alpha1](https://huggingface.co/llm-jp/llm-jp-3-172b-alpha1), [**172b**-alpha1-instruct](https://huggingface.co/llm-jp/llm-jp-3-172b-alpha1-instruct), [**172b**-alpha2](https://huggingface.co/llm-jp/llm-jp-3-172b-alpha2), [**172b**-alpha2-instruct](https://huggingface.co/llm-jp/llm-jp-3-172b-alpha2-instruct)) | 4,096 | 事前学習: [llm-jp-corpus-v3](https://gitlab.llm-jp.nii.ac.jp/datasets/llm-jp-corpus-v3)の一部<br>(alpha1: 計 **0.7T** トークン, alpha2: 計 **1.4T** トークン)<br>Instruction Tuning: [ichikara-instruction](https://liat-aip.sakura.ne.jp/wp/llm%E3%81%AE%E3%81%9F%E3%82%81%E3%81%AE%E6%97%A5%E6%9C%AC%E8%AA%9E%E3%82%A4%E3%83%B3%E3%82%B9%E3%83%88%E3%83%A9%E3%82%AF%E3%82%B7%E3%83%A7%E3%83%B3%E3%83%87%E3%83%BC%E3%82%BF%E4%BD%9C%E6%88%90/), [answer-carefully](https://liat-aip.sakura.ne.jp/wp/answercarefully-dataset/), Dolly Dataset, OASST1, OASST2, Aya Dataset, 	ichikara-instruction-format, Daring-Anteater, FLAN | 大規模言語モデル研究開発センター (LLMC) | Apache 2.0 |
 | [Stockmark-100b](https://stockmark.co.jp/news/20240516) | Llama<br>([**100b**](https://huggingface.co/stockmark/stockmark-100b), [**100b**-instruct-v0.1](https://huggingface.co/stockmark/stockmark-100b-instruct-v0.1)) | 4,096 | 事前学習: RedPajama, 日本語 Wikipedia, Japanese mC4, Japanese CommonCrawl, 日本語特許, Stockmark Web Corpus<br>(計 **910B** トークン)<br>Instruction Tuning (LoRA): [ichikara-instruction](https://liat-aip.sakura.ne.jp/wp/llm%E3%81%AE%E3%81%9F%E3%82%81%E3%81%AE%E6%97%A5%E6%9C%AC%E8%AA%9E%E3%82%A4%E3%83%B3%E3%82%B9%E3%83%88%E3%83%A9%E3%82%AF%E3%82%B7%E3%83%A7%E3%83%B3%E3%83%87%E3%83%BC%E3%82%BF%E4%BD%9C%E6%88%90/) | ストックマーク | MIT |
+| [PLaMo-100B-Pretrained](https://www.preferred.jp/ja/news/pr20241015/) | Llama[^22]<br>([**100b**](https://huggingface.co/pfnet/plamo-100b)) | 4,096 | 事前学習: Japanese CommonCrawl, RefinedWeb, 独自のデータセット<br>(計: **2.0T** トークン) | Preferred Elements | PLaMo Non-Commercial License |
 | [Sarashina2](https://www.sbintuitions.co.jp/news/press/20240614_01/) | Llama<br>([**7b**](https://huggingface.co/sbintuitions/sarashina2-7b), [**13b**](https://huggingface.co/sbintuitions/sarashina2-13b), [**70b**](https://huggingface.co/sbintuitions/sarashina2-70b)) | 7b, 13b: 4,096<br>70b: 8,192 | 事前学習: Japanese Common Crawl, SlimPajama, StarCoder<br>(計 **2.1T** トークン) | SB Intuitions | MIT |
 | [Sarashina1](https://www.sbintuitions.co.jp/news/press/20240614_01/) | GPT-NeoX<br>([**7b**](https://huggingface.co/sbintuitions/sarashina1-7b), [**13b**](https://huggingface.co/sbintuitions/sarashina1-13b), [**65b**](https://huggingface.co/sbintuitions/sarashina1-65b)) | 2,048 | 事前学習: Japanese Common Crawl<br>(計 **1T** トークン) | SB Intuitions | MIT |
 | [Tanuki-8×8B](https://weblab.t.u-tokyo.ac.jp/2024-08-30/) | Tanuki (MoE) (**47b**)<br>([v1.0](https://huggingface.co/weblab-GENIAC/Tanuki-8x8B-dpo-v1.0), [v1.0-AWQ](https://huggingface.co/team-hatakeyama-phase2/Tanuki-8x8B-dpo-v1.0-AWQ), [v1.0-GPTQ-4bit](https://huggingface.co/team-hatakeyama-phase2/Tanuki-8x8B-dpo-v1.0-GPTQ-4bit), [v1.0-GPTQ-8bit](https://huggingface.co/team-hatakeyama-phase2/Tanuki-8x8B-dpo-v1.0-GPTQ-8bit), [v1.0-GGUF](https://huggingface.co/team-hatakeyama-phase2/Tanuki-8x8B-dpo-v1.0-GGUF)) | 4,096 | 事前学習: 様々な Web 上のデータ, 合成データ（計 **1.7T** トークン）<br>SFT, DPO: 様々な合成データ [^19] | 松尾研LLM開発プロジェクト | Apache 2.0 |
@@ -562,4 +563,6 @@
 
 [^20]: ORPO を行う前に、Gemma 2 Instruct と Gemma 2 Base の差分の Chat Vector を加えている。
 
-[^21]: 埋め込みモデルの分類は [Dense Text Retrieval based on Pretrained Language Models: A Survey (Zhao+, 2022)](https://arxiv.org/abs/2211.14876) を参考に行った。Bi-Encoder は 2つの入力を個別にモデルに入力し、それぞれベクトル化した上で、それらの内積やコサイン類似度を入力の近さとして定式化するアーキテクチャである。それに対し、Cross-Encoder は 2 つの入力を組み合わせたものをモデルに入力し、モデル内部で近さを直接計算するアーキテクチャである。情報抽出の分野では、Cross-Encoder の方が計算コストがかかるが、入力の近さをよりきめ細かくモデルが計算することが期待されるため、抽出結果の順序を再検討するリランカーとして用いられることも多い。なお、Bi-Encoder の中でも、入力を単一のベクトルではなく（トークンごとなどの）複数のベクトルとして表現するタイプのもの（例: ColBERT）があるため、Single-representation bi-encoders と Multi-representation bi-encoders にさらに細分化している。
+[^21]: 埋め込みモデルの分類は [Dense Text Retrieval based on Pretrained Language Models: A Survey (Zhao+, 2022)](https://arxiv.org/abs/2211.14876) を参考に行った。Bi-Encoder は 2つの入力を個別にモデルに入力し、それぞれベクトル化した上で、それらの内積やコサイン類似度を入力の近さとして定式化するアーキテクチャである。それに対し、Cross-Encoder は 2 つの入力を組み合わせたものをモデルに入力し、モデル内部で近さを直接計算するアーキテクチャである。情報抽出の分野では、Cross-Encoder の方が計算コストがかかるが、入力の近さをよりきめ細かくモデルが計算することが期待されるため、抽出結果の順序を再検討するリランカーとして用いられることも多い。なお、Bi-Encoder の中でも、入力を単一のベクトルではなく（トークンごとなどの）複数のベクトルとして表現するタイプのもの（例: ColBERT）があるため、Single-representation bi-encoders と Multi-representation bi-encoders にさらに細分化している。
+
+[^22]: 一部アーキテクチャの変更を加えている。詳しくは以下を参照: [1,000億パラメータ規模の独自LLM「PLaMo-100B」の事前学習](https://tech.preferred.jp/ja/blog/plamo-100b/)
diff --git a/en/README.md b/en/README.md
@@ -38,6 +38,7 @@ Please point out any errors on the [issues page](https://github.com/llm-jp/aweso
 | [LLM-jp-3 172B beta1](https://www.nii.ac.jp/en/news/release/2024/0917.html) | Llama<br>([**172b**-beta1](https://huggingface.co/llm-jp/llm-jp-3-172b-beta1), [**172b**-beta1-instruct](https://huggingface.co/llm-jp/llm-jp-3-172b-beta1-instruct)) | 4,096 | Pre-training: part of [llm-jp-corpus-v3](https://gitlab.llm-jp.nii.ac.jp/datasets/llm-jp-corpus-v3)<br>(**0.7T** tokens)<br>Instruction Tuning: [ichikara-instruction](https://liat-aip.sakura.ne.jp/wp/llm%E3%81%AE%E3%81%9F%E3%82%81%E3%81%AE%E6%97%A5%E6%9C%AC%E8%AA%9E%E3%82%A4%E3%83%B3%E3%82%B9%E3%83%88%E3%83%A9%E3%82%AF%E3%82%B7%E3%83%A7%E3%83%B3%E3%83%87%E3%83%BC%E3%82%BF%E4%BD%9C%E6%88%90/), [answer-carefully](https://liat-aip.sakura.ne.jp/wp/answercarefully-dataset/), Dolly Dataset, OASST1, OASST2, Aya Dataset, 	ichikara-instruction-format, Daring-Anteater, FLAN | Research and Development Center for Large Language Models (LLMC) | LLM-jp-3 172B beta1 Terms of Use |
 | [LLM-jp-3 172B alpha](https://llmc.nii.ac.jp/en/topics/llm-jp-3-172b-alpha1-alpha2/) | Llama<br>([**172b**-alpha1](https://huggingface.co/llm-jp/llm-jp-3-172b-alpha1), [**172b**-alpha1-instruct](https://huggingface.co/llm-jp/llm-jp-3-172b-alpha1-instruct), [**172b**-alpha2](https://huggingface.co/llm-jp/llm-jp-3-172b-alpha2), [**172b**-alpha2-instruct](https://huggingface.co/llm-jp/llm-jp-3-172b-alpha2-instruct)) | 4,096 | Pre-training: part of [llm-jp-corpus-v3](https://gitlab.llm-jp.nii.ac.jp/datasets/llm-jp-corpus-v3)<br>(alpha1: **0.7T** tokens, alpha2: **1.4T** tokens)<br>Instruction Tuning: [ichikara-instruction](https://liat-aip.sakura.ne.jp/wp/llm%E3%81%AE%E3%81%9F%E3%82%81%E3%81%AE%E6%97%A5%E6%9C%AC%E8%AA%9E%E3%82%A4%E3%83%B3%E3%82%B9%E3%83%88%E3%83%A9%E3%82%AF%E3%82%B7%E3%83%A7%E3%83%B3%E3%83%87%E3%83%BC%E3%82%BF%E4%BD%9C%E6%88%90/), [answer-carefully](https://liat-aip.sakura.ne.jp/wp/answercarefully-dataset/), Dolly Dataset, OASST1, OASST2, Aya Dataset, 	ichikara-instruction-format, Daring-Anteater, FLAN | Research and Development Center for Large Language Models (LLMC) | Apache 2.0 |
 | [Stockmark-100b](https://huggingface.co/stockmark/stockmark-100b) | Llama<br>([**100b**](https://huggingface.co/stockmark/stockmark-100b), [**100b**-instruct-v0.1](https://huggingface.co/stockmark/stockmark-100b-instruct-v0.1)) | 4,096 | Pre-training: RedPajama, Japanese Wikipedia, Japanese mC4, Japanese CommonCrawl, Japanese Patent, Stockmark Web Corpus<br>(**910B** tokens)<br>Instruction Tuning (LoRA): [ichikara-instruction](https://liat-aip.sakura.ne.jp/wp/llm%E3%81%AE%E3%81%9F%E3%82%81%E3%81%AE%E6%97%A5%E6%9C%AC%E8%AA%9E%E3%82%A4%E3%83%B3%E3%82%B9%E3%83%88%E3%83%A9%E3%82%AF%E3%82%B7%E3%83%A7%E3%83%B3%E3%83%87%E3%83%BC%E3%82%BF%E4%BD%9C%E6%88%90/) | Stockmark | MIT |
+| [PLaMo-100B-Pretrained](https://www.preferred.jp/ja/news/pr20241015/) | Llama[^22]<br>([**100b**](https://huggingface.co/pfnet/plamo-100b)) | 4,096 | Pre-training: Japanese CommonCrawl, RefinedWeb, undisclosed<br>(**2.0T** tokens) | Preferred Elements | PLaMo Non-Commercial License |
 | [Sarashina2](https://www.sbintuitions.co.jp/news/press/20240614_01/) | Llama<br>([**7b**](https://huggingface.co/sbintuitions/sarashina2-7b), [**13b**](https://huggingface.co/sbintuitions/sarashina2-13b), [**70b**](https://huggingface.co/sbintuitions/sarashina2-70b)) | 7b, 13b: 4,096<br>70b: 8,192 | Pre-training: Japanese Common Crawl, SlimPajama, StarCoder<br>(**2.1T** tokens) | SB Intuitions | MIT |
 | [Sarashina1](https://www.sbintuitions.co.jp/news/press/20240614_01/) | GPT-NeoX<br>([**7b**](https://huggingface.co/sbintuitions/sarashina1-7b), [**13b**](https://huggingface.co/sbintuitions/sarashina1-13b), [**65b**](https://huggingface.co/sbintuitions/sarashina1-65b)) | 2,048 | Pre-training: Japanese Common Crawl<br>(**1T** tokens) | SB Intuitions | MIT |
 | [Tanuki-8×8B](https://weblab.t.u-tokyo.ac.jp/2024-08-30/) | Tanuki (MoE) (**47b**)<br>([v1.0](https://huggingface.co/weblab-GENIAC/Tanuki-8x8B-dpo-v1.0), [v1.0-AWQ](https://huggingface.co/team-hatakeyama-phase2/Tanuki-8x8B-dpo-v1.0-AWQ), [v1.0-GPTQ-4bit](https://huggingface.co/team-hatakeyama-phase2/Tanuki-8x8B-dpo-v1.0-GPTQ-4bit), [v1.0-GPTQ-8bit](https://huggingface.co/team-hatakeyama-phase2/Tanuki-8x8B-dpo-v1.0-GPTQ-8bit), [v1.0-GGUF](https://huggingface.co/team-hatakeyama-phase2/Tanuki-8x8B-dpo-v1.0-GGUF)) | 4,096 | Pre-training: various Web & synthetic datasets（**1.7T** tokens）<br>SFT, DPO: various synthetic datasets [^19] | Matsuo Lab LLM Development Project | Apache 2.0 |
@@ -560,4 +561,6 @@ When referencing this repository, please cite as follows:
 
 [^20]: Before conducting Instruction Tuning, a Chat Vector between Gemma 2 Instruct and Gemma 2 Base is added.
 
-[^21]: The classification of embedding models was referenced from [Dense Text Retrieval based on Pretrained Language Models: A Survey (Zhao+, 2022)](https://arxiv.org/abs/2211.14876). The Bi-Encoder architecture inputs two separate inputs into the model and vectorizes each, using their dot product or cosine similarity as a measure of their proximity. In contrast, the Cross-Encoder architecture inputs the combined inputs into the model to directly compute their proximity internally. Although Cross-Encoders incur higher computational costs, they are often used as rerankers in information extraction due to their ability to compute input proximity more precisely. Among Bi-Encoders, there are types (e.g., ColBERT) that represent the input as multiple vectors (such as one per token) rather than a single vector, hence further classification into Single-representation bi-encoders and Multi-representation bi-encoders.
+[^21]: The classification of embedding models was referenced from [Dense Text Retrieval based on Pretrained Language Models: A Survey (Zhao+, 2022)](https://arxiv.org/abs/2211.14876). The Bi-Encoder architecture inputs two separate inputs into the model and vectorizes each, using their dot product or cosine similarity as a measure of their proximity. In contrast, the Cross-Encoder architecture inputs the combined inputs into the model to directly compute their proximity internally. Although Cross-Encoders incur higher computational costs, they are often used as rerankers in information extraction due to their ability to compute input proximity more precisely. Among Bi-Encoders, there are types (e.g., ColBERT) that represent the input as multiple vectors (such as one per token) rather than a single vector, hence further classification into Single-representation bi-encoders and Multi-representation bi-encoders.
+
+[^22]: Some architectural changes have been made. For details, refer to: [1,000億パラメータ規模の独自LLM「PLaMo-100B」の事前学習](https://tech.preferred.jp/ja/blog/plamo-100b/)
diff --git a/figures/parameter_size_overview_en.png b/figures/parameter_size_overview_en.png
diff --git a/figures/parameter_size_overview_ja.png b/figures/parameter_size_overview_ja.png
diff --git a/figures/scripts/parameter_size_overview_ja.csv b/figures/scripts/parameter_size_overview_ja.csv
@@ -2,7 +2,7 @@ Model,Lab,Parameters(B),Announced,Type,Source(JP)
 日本語版  Gemma 2 2B,Google,2,2024/10/3,JP-available-CP,https://developers-jp.googleblog.com/2024/10/gemma-2-for-japan.html
 Sarashina2-70b,SB Intuitions,70,2024/8/7,JP-available,https://huggingface.co/sbintuitions/sarashina2-70b
 Sarashina,SB Intuitions,65,2024/6/14,JP-available,https://www.sbintuitions.co.jp/news/press/20240614_01/
-Takane,Fujitsu,104,2024/9/30,JP-available-CP,https://pr.fujitsu.com/jp/news/2024/09/30.html
+Takane,Fujitsu,104,2024/9/30,JP-unavailable,https://pr.fujitsu.com/jp/news/2024/09/30.html
 Fugaku-LLM,"Titech, Tohoku Univ., Fujitsu, RIKEN, Nagoya Univ., CyberAgent, Kotoba Technologies",13,2024/5/10,JP-available,https://www.fujitsu.com/global/about/resources/news/press-releases/2024/0510-01.html
 ,Rakuten,7,2024/3/21,JP-available-CP,https://corp.rakuten.co.jp/news/press/2024/0321_01.html
 KARAKURI 8x7B Instruct,KARAKURI,46.7,2024/6/20,JP-available-CP,https://karakuri.ai/seminar/news/karakuri-lm-8x7b-instruct-v0-1/
@@ -20,7 +20,7 @@ LLM-jp-3 13B,LLMC,13,2024/09/25,JP-available,https://llmc.nii.ac.jp/topics/post-
 LLM-jp-3 172B beta1,LLMC,172,2024/09/17,JP-available,https://www.nii.ac.jp/news/release/2024/0917.html
 LLM-jp-13B v2.0,LLM-jp,13,2024/04/30,JP-available,https://www.nii.ac.jp/news/release/2024/0430.html
 LLM-jp-13B,LLM-jp,13,2023/10/20,JP-available,https://www.nii.ac.jp/news/release/2023/1020.html
-PLaMo-100B,Preferred Elements,100,2024/06/14,JP-unavailable,https://tech.preferred.jp/ja/blog/plamo-100b/
+PLaMo-100B,Preferred Elements,100,2024/06/14,JP-available,"https://tech.preferred.jp/ja/blog/plamo-100b/, https://www.preferred.jp/ja/news/pr20241015/"
 PLaMo-13B,Preferred Networks,13,2023/09/28,JP-available,https://www.preferred.jp/ja/news/pr20230928/
 Stockmark-100b,Stockmark,100,2024/05/16,JP-available,https://stockmark.co.jp/news/20240516
 Stockmark-13b,Stockmark,13,2023/10/27,JP-available,https://stockmark.co.jp/news/20231027