Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,256 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: other
|
3 |
+
license_name: llama-3
|
4 |
+
license_link: https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct/raw/main/LICENSE
|
5 |
+
|
6 |
+
base_model: meta-llama/Meta-Llama-3-8B-Instruct
|
7 |
+
tags:
|
8 |
+
- generated_from_trainer
|
9 |
+
model-index:
|
10 |
+
- name: lightblue/suzume-llama-3-8B-multilingual
|
11 |
+
results: []
|
12 |
+
---
|
13 |
+
|
14 |
+
<p align="center">
|
15 |
+
<img width=400 src="https://cdn-uploads.huggingface.co/production/uploads/64b63f8ad57e02621dc93c8b/kg3QjQOde0X743csGJT-f.png" alt="Suzume - a Japanese tree sparrow"/>
|
16 |
+
</p>
|
17 |
+
|
18 |
+
# Suzume
|
19 |
+
|
20 |
+
This Suzume 8B, a multilingual finetune of Llama 3.
|
21 |
+
|
22 |
+
Llama 3 has exhibited excellent performance on many English language benchmarks.
|
23 |
+
However, it also seemingly been finetuned on mostly English data, meaning that it will respond in English, even if prompted in other languages.
|
24 |
+
|
25 |
+
We have fine-tuned Llama 3 on almost 90,000 multilingual conversations meaning that this model has the smarts of Llama 3 but has the added ability to chat in more languages.
|
26 |
+
|
27 |
+
Please feel free to comment on this model and give us feedback in the Community tab!
|
28 |
+
|
29 |
+
# How to use
|
30 |
+
|
31 |
+
The easiest way to use this model on your own computer is to use the [GGUF version of this model (lightblue/suzume-llama-3-8B-multilingual-gguf)](https://huggingface.co/lightblue/suzume-llama-3-8B-multilingual-gguf) using a program such as (jan.ai)[https://jan.ai/] or [LM Studio](https://lmstudio.ai/).
|
32 |
+
|
33 |
+
If you want to use this model directly in Python, we recommend using vLLM for the fastest inference speeds.
|
34 |
+
|
35 |
+
```python
|
36 |
+
from vllm import LLM, SamplingParams
|
37 |
+
|
38 |
+
sampling_params = SamplingParams(temperature=0.0, max_tokens=100)
|
39 |
+
llm = LLM(model="lightblue/suzume-llama-3-8B-multilingual")
|
40 |
+
|
41 |
+
messages = []
|
42 |
+
messages.append({"role": "user", "content": "Bonjour!"})
|
43 |
+
prompt = llm.llm_engine.tokenizer.tokenizer.apply_chat_template(conversation=messages, add_generation_prompt=True, tokenize=False)
|
44 |
+
prompts = [prompt]
|
45 |
+
|
46 |
+
outputs = llm.generate(prompts, sampling_params)
|
47 |
+
for output in outputs:
|
48 |
+
prompt = output.prompt
|
49 |
+
generated_text = output.outputs[0].text
|
50 |
+
print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
|
51 |
+
```
|
52 |
+
|
53 |
+
# Evaluation scores
|
54 |
+
|
55 |
+
We achieve the following MT-Bench scores across 6 languages:
|
56 |
+
|
57 |
+
| | **meta-llama/Meta-Llama-3-8B-Instruct** | **lightblue/suzume-llama-3-8B-multilingual** | **Nexusflow/Starling-LM-7B-beta** | **gpt-3.5-turbo** |
|
58 |
+
|-----------------|-----------------------------------------|----------------------------------------------|-----------------------------------|-------------------|
|
59 |
+
| **German** π©πͺ | NaN | 7.26 | 6.99 | 7.68 |
|
60 |
+
| **French** π«π· | NaN | 7.66 | 7.29 | 7.74 |
|
61 |
+
| **Japanese** π―π΅ | NaN | 6.56 | 6.22 | 7.84 |
|
62 |
+
| **Russian** π·πΊ | NaN | 8.19 | 8.28 | 7.94 |
|
63 |
+
| **Chinese** π¨π³ | NaN | 7.11 | 6.97 | 7.55 |
|
64 |
+
| **English** πΊπΈ | 7.98 | 7.73 | 7.92 | 8.26 |
|
65 |
+
|
66 |
+
We observe minimal degredation of Llama 3's English ability while achieving best-in-class multilingual abilities compared to the top rated 7B model ([Nexusflow/Starling-LM-7B-beta](https://huggingface.co/Nexusflow/Starling-LM-7B-beta)) on the [Chatbot Arena Leaderboard](https://chat.lmsys.org/?leaderboard).
|
67 |
+
|
68 |
+
[Here is our evaluation script.](https://drive.google.com/file/d/15HPn7452t8LbTD9HKSl7ngYYWnsoOG08/view?usp=sharing)
|
69 |
+
|
70 |
+
# Training data
|
71 |
+
|
72 |
+
We train on three sources of data to create this model:
|
73 |
+
|
74 |
+
* [lightblue/tagengo-gpt4](https://huggingface.co/datasets/lightblue/tagengo-gpt4) - 76,338 conversations
|
75 |
+
* A diverse dataset of initial inputs sampled from [lmsys/lmsys-chat-1m](https://huggingface.co/datasets/lmsys/lmsys-chat-1m) and then used to prompt `gpt-4-0125-preview`
|
76 |
+
* [megagonlabs/instruction_ja](https://github.com/megagonlabs/instruction_ja) - 669 conversations
|
77 |
+
* A hand-edited dataset of nearly 700 Japanese conversations taken originally from translations of the [kunishou/hh-rlhf-49k-ja](https://huggingface.co/datasets/kunishou/hh-rlhf-49k-ja) dataset.
|
78 |
+
* [openchat/openchat_sharegpt4_dataset](https://huggingface.co/datasets/openchat/openchat_sharegpt4_dataset/resolve/main/sharegpt_gpt4.json) - 6,206 conversations
|
79 |
+
* Multilingual conversations of humans talking to GPT-4.
|
80 |
+
|
81 |
+
|
82 |
+
<details><summary>We prepare our data like so:</summary>
|
83 |
+
|
84 |
+
```python
|
85 |
+
import pandas as pd
|
86 |
+
from datasets import Dataset, load_dataset, concatenate_datasets
|
87 |
+
|
88 |
+
### Tagengo
|
89 |
+
gpt4_dataset = load_dataset("lightblue/tagengo-gpt4", split="train")
|
90 |
+
gpt4_dataset = gpt4_dataset.filter(lambda x: x["response"][1] == "stop")
|
91 |
+
####
|
92 |
+
|
93 |
+
### Megagon
|
94 |
+
megagon_df = pd.read_json(
|
95 |
+
"https://raw.githubusercontent.com/megagonlabs/instruction_ja/main/data/data.jsonl",
|
96 |
+
lines=True,
|
97 |
+
orient="records"
|
98 |
+
)
|
99 |
+
role_map = {"user": "human", "agent": "gpt"}
|
100 |
+
megagon_df["conversations"] = megagon_df.utterances.apply(lambda x: [{"from": role_map[y["name"]], "value": y["text"]} for y in x])
|
101 |
+
megagon_df["language"] = "Japanese"
|
102 |
+
megagon_df = megagon_df[["conversations", "language"]]
|
103 |
+
megagon_dataset = Dataset.from_pandas(df)
|
104 |
+
###
|
105 |
+
|
106 |
+
### Openchat
|
107 |
+
openchat_df = pd.read_json("https://huggingface.co/datasets/openchat/openchat_sharegpt4_dataset/resolve/main/sharegpt_gpt4.json?download=true")
|
108 |
+
openchat_df["conversations"] = openchat_df["items"]
|
109 |
+
openchat_dataset = Dataset.from_pandas(openchat_df)
|
110 |
+
###
|
111 |
+
|
112 |
+
|
113 |
+
dataset = concatenate_datasets([gpt4_dataset, megagon_dataset, openchat_dataset])
|
114 |
+
dataset = dataset.filter(lambda x: not any([y["value"] is None for y in x["conversations"]]))
|
115 |
+
dataset.select_columns(["conversations"]).to_json("/workspace/llm_training/axolotl/llama3-multilingual/tagengo_openchat_megagon.json")
|
116 |
+
```
|
117 |
+
|
118 |
+
</details>
|
119 |
+
<br/>
|
120 |
+
|
121 |
+
# workspace/llm_training/axolotl/llama3-multilingual/output_tagengo_openchat_megagon_8B_llama3
|
122 |
+
|
123 |
+
This model is a fine-tuned version of [meta-llama/Meta-Llama-3-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct) on the above described dataset.
|
124 |
+
It achieves the following results on the evaluation set:
|
125 |
+
- Loss: 0.6595
|
126 |
+
|
127 |
+
|
128 |
+
## Training procedure
|
129 |
+
|
130 |
+
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
|
131 |
+
should probably proofread and complete it, then remove this comment. -->
|
132 |
+
|
133 |
+
[<img src="https://raw.githubusercontent.com/OpenAccess-AI-Collective/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/OpenAccess-AI-Collective/axolotl)
|
134 |
+
<details><summary>See axolotl config</summary>
|
135 |
+
|
136 |
+
axolotl version: `0.4.0`
|
137 |
+
```yaml
|
138 |
+
base_model: meta-llama/Meta-Llama-3-8B-Instruct
|
139 |
+
model_type: LlamaForCausalLM
|
140 |
+
tokenizer_type: AutoTokenizer # PreTrainedTokenizerFast
|
141 |
+
|
142 |
+
load_in_8bit: false
|
143 |
+
load_in_4bit: false
|
144 |
+
strict: false
|
145 |
+
|
146 |
+
datasets:
|
147 |
+
- path: /workspace/llm_training/axolotl/llama3-multilingual/tagengo_openchat_megagon.json
|
148 |
+
ds_type: json # see other options below
|
149 |
+
type: sharegpt
|
150 |
+
conversation: llama-3
|
151 |
+
dataset_prepared_path: /workspace/llm_training/axolotl/llama3-multilingual/prepared_tagengo_openchat_megagon
|
152 |
+
val_set_size: 0.01
|
153 |
+
output_dir: /workspace/llm_training/axolotl/llama3-multilingual/output_tagengo_openchat_megagon_8B_llama3
|
154 |
+
|
155 |
+
sequence_len: 8192
|
156 |
+
sample_packing: true
|
157 |
+
pad_to_sequence_len: true
|
158 |
+
|
159 |
+
use_wandb: true
|
160 |
+
wandb_project: wandb_project
|
161 |
+
wandb_entity: wandb_entity
|
162 |
+
wandb_name: wandb_name
|
163 |
+
|
164 |
+
gradient_accumulation_steps: 2
|
165 |
+
micro_batch_size: 2
|
166 |
+
num_epochs: 1
|
167 |
+
optimizer: paged_adamw_8bit
|
168 |
+
lr_scheduler: cosine
|
169 |
+
learning_rate: 1e-5
|
170 |
+
|
171 |
+
train_on_inputs: false
|
172 |
+
group_by_length: false
|
173 |
+
bf16: auto
|
174 |
+
fp16:
|
175 |
+
tf32: false
|
176 |
+
|
177 |
+
gradient_checkpointing: true
|
178 |
+
gradient_checkpointing_kwargs:
|
179 |
+
use_reentrant: false
|
180 |
+
early_stopping_patience:
|
181 |
+
resume_from_checkpoint:
|
182 |
+
logging_steps: 1
|
183 |
+
xformers_attention:
|
184 |
+
flash_attention: true
|
185 |
+
|
186 |
+
warmup_steps: 10
|
187 |
+
evals_per_epoch: 5
|
188 |
+
eval_table_size:
|
189 |
+
saves_per_epoch: 1
|
190 |
+
debug:
|
191 |
+
deepspeed: /workspace/axolotl/deepspeed_configs/zero2.json
|
192 |
+
weight_decay: 0.0
|
193 |
+
special_tokens:
|
194 |
+
pad_token: <|end_of_text|>
|
195 |
+
```
|
196 |
+
|
197 |
+
</details><br>
|
198 |
+
|
199 |
+
<details><summary>Note - we added this Llama 3 template to fastchat directly as the Llama 3 chat template was not supported when we trained this model.</summary>
|
200 |
+
|
201 |
+
```python
|
202 |
+
from fastchat.conversation import Conversation
|
203 |
+
from fastchat.conversation import register_conv_template
|
204 |
+
from fastchat.conversation import SeparatorStyle
|
205 |
+
|
206 |
+
register_conv_template(
|
207 |
+
Conversation(
|
208 |
+
name="llama-3",
|
209 |
+
system_template="<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n{system_message}",
|
210 |
+
roles=("<|start_header_id|>user<|end_header_id|>\n", "<|start_header_id|>assistant<|end_header_id|>\n"),
|
211 |
+
sep_style=SeparatorStyle.ADD_NEW_LINE_SINGLE,
|
212 |
+
sep="<|eot_id|>",
|
213 |
+
stop_token_ids=[128009],
|
214 |
+
stop_str="<|eot_id|>",
|
215 |
+
)
|
216 |
+
)
|
217 |
+
```
|
218 |
+
|
219 |
+
</details><br>
|
220 |
+
|
221 |
+
|
222 |
+
### Training hyperparameters
|
223 |
+
|
224 |
+
The following hyperparameters were used during training:
|
225 |
+
- learning_rate: 1e-05
|
226 |
+
- train_batch_size: 2
|
227 |
+
- eval_batch_size: 2
|
228 |
+
- seed: 42
|
229 |
+
- distributed_type: multi-GPU
|
230 |
+
- num_devices: 4
|
231 |
+
- gradient_accumulation_steps: 2
|
232 |
+
- total_train_batch_size: 16
|
233 |
+
- total_eval_batch_size: 8
|
234 |
+
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
|
235 |
+
- lr_scheduler_type: cosine
|
236 |
+
- lr_scheduler_warmup_steps: 10
|
237 |
+
- num_epochs: 1
|
238 |
+
|
239 |
+
### Training results
|
240 |
+
|
241 |
+
| Training Loss | Epoch | Step | Validation Loss |
|
242 |
+
|:-------------:|:-----:|:----:|:---------------:|
|
243 |
+
| 1.1894 | 0.0 | 1 | 1.0110 |
|
244 |
+
| 0.8493 | 0.2 | 73 | 0.7057 |
|
245 |
+
| 0.8047 | 0.4 | 146 | 0.6835 |
|
246 |
+
| 0.7644 | 0.6 | 219 | 0.6687 |
|
247 |
+
| 0.7528 | 0.8 | 292 | 0.6615 |
|
248 |
+
| 0.7794 | 1.0 | 365 | 0.6595 |
|
249 |
+
|
250 |
+
|
251 |
+
### Framework versions
|
252 |
+
|
253 |
+
- Transformers 4.38.2
|
254 |
+
- Pytorch 2.2.1+cu121
|
255 |
+
- Datasets 2.18.0
|
256 |
+
- Tokenizers 0.15.0
|