Text Generation
Transformers
Safetensors
Japanese
llama
tiny
japanese
character
zundamon
llama2.c
esp32
text-generation-inference
Instructions to use k4rasu/zundamon-tiny-llm with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use k4rasu/zundamon-tiny-llm with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="k4rasu/zundamon-tiny-llm")# pip install -U transformers accelerate # Load model directly from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("k4rasu/zundamon-tiny-llm") model = AutoModelForCausalLM.from_pretrained("k4rasu/zundamon-tiny-llm", device_map="auto") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use k4rasu/zundamon-tiny-llm with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "k4rasu/zundamon-tiny-llm" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "k4rasu/zundamon-tiny-llm", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker
docker model run hf.co/k4rasu/zundamon-tiny-llm
- SGLang
How to use k4rasu/zundamon-tiny-llm with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "k4rasu/zundamon-tiny-llm" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "k4rasu/zundamon-tiny-llm", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "k4rasu/zundamon-tiny-llm" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "k4rasu/zundamon-tiny-llm", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }' - Docker Model Runner
How to use k4rasu/zundamon-tiny-llm with Docker Model Runner:
docker model run hf.co/k4rasu/zundamon-tiny-llm
Upload folder using huggingface_hub
Browse files- README.md +116 -0
- config.json +33 -0
- generation_config.json +19 -0
- llama2c/tok.bin +3 -0
- llama2c/zundamon-v4-base_fp32.bin +3 -0
- llama2c/zundamon-v4-base_q8.bin +3 -0
- llama2c/zundamon-v4-small_fp32.bin +3 -0
- llama2c/zundamon-v4-small_q8.bin +3 -0
- model.safetensors +3 -0
- small/config.json +33 -0
- small/generation_config.json +19 -0
- small/model.safetensors +3 -0
- small/tokenizer.json +0 -0
- small/tokenizer_config.json +9 -0
- tokenizer.json +0 -0
- tokenizer_config.json +9 -0
README.md
ADDED
|
@@ -0,0 +1,116 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- ja
|
| 4 |
+
license: mit
|
| 5 |
+
library_name: transformers
|
| 6 |
+
pipeline_tag: text-generation
|
| 7 |
+
tags:
|
| 8 |
+
- llama
|
| 9 |
+
- tiny
|
| 10 |
+
- japanese
|
| 11 |
+
- character
|
| 12 |
+
- zundamon
|
| 13 |
+
- llama2.c
|
| 14 |
+
- esp32
|
| 15 |
+
---
|
| 16 |
+
|
| 17 |
+
# 幼生ずんだもん v4(zundamon-tiny-llm)
|
| 18 |
+
|
| 19 |
+
ゼロから学習した、とても小さな「ずんだもん風」の日本語会話モデルです。話しかけると、ずんだもんとして短く返事をし、返事の最後に、そのときの気持ちのタグ(〈うれしい〉〈かなしい〉など 8 種類)を 1 つ付けます。
|
| 20 |
+
|
| 21 |
+
| 版 | パラメータ | 構造 | int8 の大きさ |
|
| 22 |
+
|---|---|---|---|
|
| 23 |
+
| base(このフォルダの直下) | 約 174 万 | Llama(dim 128・8 層・8 ヘッド・KV 4 ヘッド・FFN 352) | 2.0MB |
|
| 24 |
+
| small(`small/`) | 約 81 万 | Llama(dim 96・6 層・6 ヘッド・KV 3 ヘッド・FFN 256) | 0.9MB |
|
| 25 |
+
|
| 26 |
+
- 語彙:2,048(SentencePiece BPE)
|
| 27 |
+
- 文脈の長さ:256 トークン
|
| 28 |
+
|
| 29 |
+
ブラウザの中だけでも動きます。M5Stack Core2(ESP32)向けの int8 版も `llama2c/` に入れています。
|
| 30 |
+
|
| 31 |
+
- **デモ**:https://crowpilot.github.io/zundamon-tiny-llm/
|
| 32 |
+
- **コード**(学習・Core2・ブラウザ):https://github.com/crowpilot/zundamon-tiny-llm
|
| 33 |
+
|
| 34 |
+
非公式のファン作品です。ずんだもんは[東北ずん子・ずんだもんプロジェクト](https://zunko.jp/)のキャラクターで、[キャラクター利用の手引き](https://zunko.jp/guideline.html)に沿って非商用で公開しています。モデルがとても小さいので、ときどきおかしなことを言います。
|
| 35 |
+
|
| 36 |
+
## 使い方(transformers)
|
| 37 |
+
|
| 38 |
+
```python
|
| 39 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 40 |
+
|
| 41 |
+
repo = "crowpilot/zundamon-tiny-llm"
|
| 42 |
+
tok = AutoTokenizer.from_pretrained(repo)
|
| 43 |
+
model = AutoModelForCausalLM.from_pretrained(repo) # small は subfolder="small"
|
| 44 |
+
|
| 45 |
+
prompt = "A:おはよう、ずんだもん!B:"
|
| 46 |
+
ids = tok(prompt, return_tensors="pt").input_ids # 先頭に <s> が付きます
|
| 47 |
+
out = model.generate(ids, max_new_tokens=48, do_sample=False)
|
| 48 |
+
print(tok.decode(out[0, ids.shape[1]:])) # 例: おはよー!ボクはいつも元気いっぱいなのだ!〈ふつう〉
|
| 49 |
+
```
|
| 50 |
+
|
| 51 |
+
### 書き方の約束
|
| 52 |
+
|
| 53 |
+
- **話者**:全角の `A:`(話しかけ)と `B:`(ずんだもん)で書きます。
|
| 54 |
+
- **返事の終わり**:返事の最後に気持ちのタグが 1 つ出て止まります(`generation_config.json` の終わりのトークンにタグを入れてあります)。8 種類は〈ふつう〉〈うれしい〉〈ドヤ〉〈びっくり〉〈かなしい〉〈ぷんぷん〉〈こまった〉〈ねむい〉です。
|
| 55 |
+
- **会話の区切り**:この版は 1 往復ずつ学習しています。話しかけるたびに先頭に `<s>` を置くと、いちばん安定します。
|
| 56 |
+
|
| 57 |
+
## Core2(ESP32)・llama2.c で使う
|
| 58 |
+
|
| 59 |
+
`llama2c/` に次のファイルがあります。
|
| 60 |
+
|
| 61 |
+
| ファイル | 中身 |
|
| 62 |
+
|---|---|
|
| 63 |
+
| `zundamon-v4-base_q8.bin`、`zundamon-v4-small_q8.bin` | int8(グループ 32)。Core2 とブラウザ版の形式 |
|
| 64 |
+
| `zundamon-v4-*_fp32.bin` | llama2.c 形式の FP32 |
|
| 65 |
+
| `tok.bin` | llama2.c 形式のトークナイザ |
|
| 66 |
+
|
| 67 |
+
読み込み方はコードのリポジトリを見てください。気持ちタグ(語彙の 5〜12 番)は、返事の終わりの印として扱います。
|
| 68 |
+
|
| 69 |
+
## 学習
|
| 70 |
+
|
| 71 |
+
- **方法**:ランダムな初期値から、次のトークンを当てる学習だけで作りました。指示に従う追加学習や強化学習はしていません。
|
| 72 |
+
- **データ**:Gemma 4(Apache 2.0)で作り、機械と Gemma の採点で選んだ、ずんだもんへの話しかけと返事の組(約 7 万組)。返事の気持ちも Gemma が 8 種類から選んでいます。学習データとデータの作り方は公開していません。
|
| 73 |
+
- **大きさ**:学習トークンは約 162 万です。
|
| 74 |
+
- **環境**:NVIDIA V100 1 枚で、base は約 10 分で学習できます。
|
| 75 |
+
|
| 76 |
+
## 評価
|
| 77 |
+
|
| 78 |
+
話しかけ 201 個への返事を、Gemma 4 がモデル名を伏せて 1〜5 点で採点しました。比べたのは、前に公開していた版(v3、気持ちタグなし)です。
|
| 79 |
+
|
| 80 |
+
| モデル | 応答 | ならでは(決まり文句でないか) | 気持ちタグの一致 |
|
| 81 |
+
|---|---|---|---|
|
| 82 |
+
| v3 base(前の公開版) | 1.89 | 1.47 | — |
|
| 83 |
+
| v4 base | 2.80 | 2.11 | 75.2% |
|
| 84 |
+
| v4 small | 2.92 | 2.07 | 75.5% |
|
| 85 |
+
|
| 86 |
+
- 「気持ちタグの一致」は、学習に使っていない組で、Gemma が付けた気持ちとモデルが選んだ気持ちが同じだった割合です。
|
| 87 |
+
- 応答の点が 2 以下の返事も、まだ半分ほどあります。
|
| 88 |
+
|
| 89 |
+
## 制限
|
| 90 |
+
|
| 91 |
+
- とても小さいモデルです。事実の知識はほとんどなく、話がつながらないこともよくあります。
|
| 92 |
+
- 1 往復ずつ学習しているので、前の会話の内容は使えません。
|
| 93 |
+
- 漢字の多い文や長い文は苦手です。
|
| 94 |
+
- 会話の練習用・遊び用です。大事な判��には使わないでください。
|
| 95 |
+
|
| 96 |
+
## ライセンス
|
| 97 |
+
|
| 98 |
+
- **モデル**:MIT
|
| 99 |
+
- **学習データ**:Gemma 4(Apache 2.0)で生成しました。
|
| 100 |
+
- **ずんだもん**:キャラクターに関する権利は東北ずん子・ずんだもんプロジェクトにあり、このライセンスの対象外です。利用するときは[キャラクター利用の手引き](https://zunko.jp/guideline.html)に従ってください。
|
| 101 |
+
|
| 102 |
+
---
|
| 103 |
+
|
| 104 |
+
## English summary
|
| 105 |
+
|
| 106 |
+
A very small Japanese chat model, trained from scratch, that answers in the style of Zundamon (unofficial, non-commercial fan work).
|
| 107 |
+
|
| 108 |
+
- **Sizes**:Llama architecture, about 1.74M parameters (base) and 0.81M (small)
|
| 109 |
+
- **Vocabulary**:2,048-token SentencePiece BPE
|
| 110 |
+
- **Context length**:256 tokens
|
| 111 |
+
|
| 112 |
+
**Prompt format**:`A:<user>B:`, with `<s>` placed before every exchange. The model ends each reply with one emotion tag such as `〈うれしい〉`, and generation stops there.
|
| 113 |
+
|
| 114 |
+
**Training data**:about 70k single-turn exchanges written by Gemma 4 (Apache 2.0) and then filtered; each reply was also labeled with one of 8 emotions by Gemma. The dataset is not released.
|
| 115 |
+
|
| 116 |
+
**Other formats**:`llama2c/` contains the int8 build for an ESP32 (M5Stack Core2) and for the browser demo.
|
config.json
ADDED
|
@@ -0,0 +1,33 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"LlamaForCausalLM"
|
| 4 |
+
],
|
| 5 |
+
"attention_bias": false,
|
| 6 |
+
"attention_dropout": 0.0,
|
| 7 |
+
"bos_token_id": 1,
|
| 8 |
+
"dtype": "float32",
|
| 9 |
+
"eos_token_id": 2,
|
| 10 |
+
"head_dim": 16,
|
| 11 |
+
"hidden_act": "silu",
|
| 12 |
+
"hidden_size": 128,
|
| 13 |
+
"initializer_range": 0.02,
|
| 14 |
+
"intermediate_size": 352,
|
| 15 |
+
"max_position_embeddings": 256,
|
| 16 |
+
"mlp_bias": false,
|
| 17 |
+
"model_type": "llama",
|
| 18 |
+
"num_attention_heads": 8,
|
| 19 |
+
"num_hidden_layers": 8,
|
| 20 |
+
"num_key_value_heads": 4,
|
| 21 |
+
"pad_token_id": null,
|
| 22 |
+
"pretraining_tp": 1,
|
| 23 |
+
"rms_norm_eps": 1e-05,
|
| 24 |
+
"rope_parameters": {
|
| 25 |
+
"rope_theta": 10000.0,
|
| 26 |
+
"rope_type": "default"
|
| 27 |
+
},
|
| 28 |
+
"tie_word_embeddings": true,
|
| 29 |
+
"transformers_version": "5.20.2",
|
| 30 |
+
"use_cache": true,
|
| 31 |
+
"vocab_size": 2048,
|
| 32 |
+
"rope_theta": 10000.0
|
| 33 |
+
}
|
generation_config.json
ADDED
|
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"bos_token_id": 1,
|
| 3 |
+
"do_sample": false,
|
| 4 |
+
"eos_token_id": [
|
| 5 |
+
2,
|
| 6 |
+
3,
|
| 7 |
+
4,
|
| 8 |
+
5,
|
| 9 |
+
6,
|
| 10 |
+
7,
|
| 11 |
+
8,
|
| 12 |
+
9,
|
| 13 |
+
10,
|
| 14 |
+
11,
|
| 15 |
+
12
|
| 16 |
+
],
|
| 17 |
+
"max_new_tokens": 48,
|
| 18 |
+
"transformers_version": "5.20.2"
|
| 19 |
+
}
|
llama2c/tok.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:213ed9474b9d447e86f901e7ad0009047c6d06887fd0fedbd8ddba5590ce2637
|
| 3 |
+
size 29754
|
llama2c/zundamon-v4-base_fp32.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f37b3daab1cbf0f8a4472a6f836aacf2c252fb1c7a1aa73b59ccc46a4d6e05bd
|
| 3 |
+
size 6971932
|
llama2c/zundamon-v4-base_q8.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c6c7553307b9705eaf6b68f41b325d43e98182b9ce44e48892c87fcb809de78d
|
| 3 |
+
size 1962536
|
llama2c/zundamon-v4-small_fp32.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ee4c89cb213b1bb872bba8bd06e3b7665688569a8404e6a9ae5bba4b7425105b
|
| 3 |
+
size 3240860
|
llama2c/zundamon-v4-small_q8.bin
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b8ef3569293a5ae8081feea9cd911fc375553ff369e3276920e51a848db001cd
|
| 3 |
+
size 910504
|
model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:303d88045dfda34d8176a2b66efbf4f97a0aafd1473e2631b75842e63a760063
|
| 3 |
+
size 6963400
|
small/config.json
ADDED
|
@@ -0,0 +1,33 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"LlamaForCausalLM"
|
| 4 |
+
],
|
| 5 |
+
"attention_bias": false,
|
| 6 |
+
"attention_dropout": 0.0,
|
| 7 |
+
"bos_token_id": 1,
|
| 8 |
+
"dtype": "float32",
|
| 9 |
+
"eos_token_id": 2,
|
| 10 |
+
"head_dim": 16,
|
| 11 |
+
"hidden_act": "silu",
|
| 12 |
+
"hidden_size": 96,
|
| 13 |
+
"initializer_range": 0.02,
|
| 14 |
+
"intermediate_size": 256,
|
| 15 |
+
"max_position_embeddings": 256,
|
| 16 |
+
"mlp_bias": false,
|
| 17 |
+
"model_type": "llama",
|
| 18 |
+
"num_attention_heads": 6,
|
| 19 |
+
"num_hidden_layers": 6,
|
| 20 |
+
"num_key_value_heads": 3,
|
| 21 |
+
"pad_token_id": null,
|
| 22 |
+
"pretraining_tp": 1,
|
| 23 |
+
"rms_norm_eps": 1e-05,
|
| 24 |
+
"rope_parameters": {
|
| 25 |
+
"rope_theta": 10000.0,
|
| 26 |
+
"rope_type": "default"
|
| 27 |
+
},
|
| 28 |
+
"tie_word_embeddings": true,
|
| 29 |
+
"transformers_version": "5.20.2",
|
| 30 |
+
"use_cache": true,
|
| 31 |
+
"vocab_size": 2048,
|
| 32 |
+
"rope_theta": 10000.0
|
| 33 |
+
}
|
small/generation_config.json
ADDED
|
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"bos_token_id": 1,
|
| 3 |
+
"do_sample": false,
|
| 4 |
+
"eos_token_id": [
|
| 5 |
+
2,
|
| 6 |
+
3,
|
| 7 |
+
4,
|
| 8 |
+
5,
|
| 9 |
+
6,
|
| 10 |
+
7,
|
| 11 |
+
8,
|
| 12 |
+
9,
|
| 13 |
+
10,
|
| 14 |
+
11,
|
| 15 |
+
12
|
| 16 |
+
],
|
| 17 |
+
"max_new_tokens": 48,
|
| 18 |
+
"transformers_version": "5.20.2"
|
| 19 |
+
}
|
small/model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:69d00b35476d01b4dc7703b5ba08baa2f5a3a0f8c981aa022e45c960898c162a
|
| 3 |
+
size 3230336
|
small/tokenizer.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
small/tokenizer_config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"bos_token": "<s>",
|
| 3 |
+
"clean_up_tokenization_spaces": false,
|
| 4 |
+
"eos_token": "</s>",
|
| 5 |
+
"model_max_length": 256,
|
| 6 |
+
"tokenizer_class": "PreTrainedTokenizerFast",
|
| 7 |
+
"unk_token": "<unk>",
|
| 8 |
+
"add_bos_token": true
|
| 9 |
+
}
|
tokenizer.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
tokenizer_config.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"bos_token": "<s>",
|
| 3 |
+
"clean_up_tokenization_spaces": false,
|
| 4 |
+
"eos_token": "</s>",
|
| 5 |
+
"model_max_length": 256,
|
| 6 |
+
"tokenizer_class": "PreTrainedTokenizerFast",
|
| 7 |
+
"unk_token": "<unk>",
|
| 8 |
+
"add_bos_token": true
|
| 9 |
+
}
|