|
--- |
|
language: |
|
- zh |
|
- en |
|
license: other |
|
tags: |
|
- llama3 |
|
- chinese |
|
- meta |
|
pipeline_tag: text-generation |
|
license_name: llama3 |
|
license_link: LICENSE |
|
--- |
|
|
|
|
|
# llama-3-8b-instruct-262k-chinese |
|
|
|
|
|
llama-3-8b-instruct-262k-chinese基于[Llama-3-8B-Instruct-262k](https://huggingface.co/gradientai/Llama-3-8B-Instruct-262k),使用ORPO方法,在中英文偏好数据集[shibing624/DPO-En-Zh-20k-Preference](https://huggingface.co/datasets/shibing624/DPO-En-Zh-20k-Preference) |
|
上微调得到的对话模型。 |
|
|
|
模型的部署、训练等方法详见MedicalGPT的GitHub仓库:[https://github.com/shibing624/MedicalGPT](https://github.com/shibing624/MedicalGPT) |
|
|
|
## Relate models |
|
- 完整模型权重:https://huggingface.co/shibing624/llama-3-8b-instruct-262k-chinese |
|
- lora权重:https://huggingface.co/shibing624/llama-3-8b-instruct-262k-chinese-lora |
|
|
|
## Features |
|
模型优势: |
|
1. 支持超长context length 262k token,适合RAG |
|
2. 支持中英文 |
|
3. 支持多轮对话,代码编码、推理能力强,英文知识充分 |
|
4. 模型推理需要显存: |
|
|
|
Quantization | Peak Usage for Encoding 2048 Tokens | Peak Usage for Generating 8192 Tokens |
|
-- | -- | -- |
|
FP16/BF16 | 18.66GB | 24.58GB |
|
Int4 | 9.21GB | 14.62GB |
|
|
|
|
|
缺点: |
|
1. model size只有8B,知识类问答幻觉明显 |
|
2. 中文知识欠缺,容易幻觉,特别是中文古文知识,属于llama类模型通病 |
|
|
|
|
|
## 如何使用 |
|
|
|
```python |
|
import transformers |
|
import torch |
|
|
|
model_id = "shibing624/llama-3-8b-instruct-262k-chinese" |
|
pipeline = transformers.pipeline( |
|
"text-generation", |
|
model=model_id, |
|
model_kwargs={"torch_dtype": torch.float16}, |
|
device="cuda", |
|
) |
|
|
|
messages = [{"role": "system", "content": ""}] |
|
messages.append({"role": "user", "content": "介绍一下机器学习"}) |
|
prompt = pipeline.tokenizer.apply_chat_template( |
|
messages, |
|
tokenize=False, |
|
add_generation_prompt=True |
|
) |
|
terminators = [ |
|
pipeline.tokenizer.eos_token_id, |
|
pipeline.tokenizer.convert_tokens_to_ids("<|eot_id|>") |
|
] |
|
outputs = pipeline( |
|
prompt, |
|
max_new_tokens=512, |
|
eos_token_id=terminators, |
|
do_sample=True, |
|
temperature=0.6, |
|
top_p=0.9 |
|
) |
|
content = outputs[0]["generated_text"][len(prompt):] |
|
print(content) |
|
``` |
|
|
|
|
|
result: |
|
```shell |
|
机器学习(Machine Learning)是一种基于计算机算法的自动数据分析技术,用于从数据中学习并预测未来的结果。它是人工智能(AI)和数据挖掘(Data Mining)的子领域,旨在通过训练和调整算法来发现数据中的模式、关系和规律。 |
|
|
|
机器学习算法可以分为监督学习、无监督学习和半监督学习三类: |
|
|
|
1. 监督学习(Supervised Learning):在这种类型的学习中,算法被提供带有标签的数据集,用于训练。算法学习如何将输入数据映射到输出数据,并在新数据上进行预测。常见的监督学习算法包括逻辑回归、决策树、支持向量机(SVM)、随机森林和神经网络。 |
|
2. 无监督学习(Unsupervised Learning):在这种类型的学习中,算法没有标签数据。算法学习数据中的模式、结构和关系,并可能发现新的数据集群或特征。常见的无监督学习算法包括聚类、主成分分析(PCA)、独立成分分析(ICA)和高维度数据降维。 |
|
3. 半监督学习(Semi-supervised Learning):在这种类型的学习中,算法被提供部分带有标签的数据集。算法学习如何将输入数据映射到输出数据,并在新数据上进行预测。半监督学习算法结合了监督学习和无监督学习的优点,常见的半监督学习算法包括自我标注(Self-Labeling)和基于图的半监督学习(Graph-based Semi-supervised Learning)。 |
|
|
|
机器学习的应用广泛,包括自然语言处理、计算机视觉、推荐系统、人工智能和自动驾驶等领域。它的优势包括: |
|
|
|
1. 自动化:机器学习算法可以自动从数据中发现模式和关系,无需人为干预。 |
|
2. 高效性:机器学习算法可以处理大量数据,并且可以在不需要人为干预的情况下进行预测。 |
|
3. 适应性:机器学习算法可以根据数据集的变化和更新进行调整。 |
|
4. 精准性:机器学习算法可以通过训练和测试来提高预测的准确性。 |
|
``` |
|
|
|
## train detail |
|
|
|
train loss: |
|
|
|
<img src="https://huggingface.co/shibing624/llama-3-8b-instruct-262k-chinese/raw/main/train_lossv2.svg" width="600"> |
|
|
|
eval loss: |
|
|
|
|
|
<img src="https://huggingface.co/shibing624/llama-3-8b-instruct-262k-chinese/raw/main/eval_lossv2.svg" width="600"> |
|
|
|
|
|
# About Llama-3-8B-Instruct-262k |
|
Gradient incorporates your data to deploy autonomous assistants that power critical operations across your business. To learn more or collaborate on a custom model. |
|
|
|
This model extends LLama-3 8B's context length from 8k to -> 160K, developed by Gradient, sponsored by compute from [Crusoe Energy](https://huggingface.co/crusoeai). It demonstrates that SOTA LLMs can learn to operate on long context with minimal training (< 200M tokens) by appropriately adjusting RoPE theta. |
|
|
|
<img src="/static-proxy?url=https%3A%2F%2Fcdn-uploads.huggingface.co%2Fproduction%2Fuploads%2F6585dc9be92bc5f258156bd6%2FhiHWva3CbsrnPvZTp5-lu.png%26quot%3B%3C%2Fspan%3E width="600"> |
|
|
|
**Approach:** |
|
|
|
- [meta-llama/Meta-Llama-3-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct) as the base |
|
- NTK-aware interpolation [1] to initialize an optimal schedule for RoPE theta, followed by a new data-driven RoPE theta optimization technique |
|
- Progressive training on increasing context lengths similar to the [Large World Model](https://huggingface.co/LargeWorldModel) [2] (See details below) |
|
|
|
**Infra:** |
|
|
|
We build on top of the EasyContext Blockwise RingAttention library [3] to scalably and efficiently train on contexts up to 262144 tokens on [Crusoe Energy](https://huggingface.co/crusoeai) high performance L40S cluster. |
|
|
|
**Data:** |
|
|
|
For training data, we generate long contexts by augmenting [SlimPajama](https://huggingface.co/datasets/cerebras/SlimPajama-627B). |
|
|
|
**Progressive Training Details:** |
|
|
|
| Parameter | 65K | 262K | |
|
|-----------------------------|----------------|------------| |
|
| Initialize From | LLaMA-3-8B-Inst| 65K | |
|
| Sequence Length | 2^16 | 2^18 | |
|
| RoPE theta | 15.3 M | 207.1 M | |
|
| Batch Size (Tokens / Step) | 2.097 M | 4.192 M | |
|
| Steps | 30 | 24 | |
|
| Total Tokens | 63 M | 101 M | |
|
| Learning Rate | 2.00E-05 | 2.00E-05 | |
|
| # GPUs | 32 | 32 | |
|
| GPU Type | NVIDIA L40S | NVIDIA L40S| |
|
|
|
|
|
|
|
|
|
|
|
|