license: other
Aquila
Aquila Language Model is the first open source language model that supports both Chinese and English knowledge, commercial license agreements, and compliance with domestic data regulations.
🌟 Supports open source commercial licenses. The source code of the Aquila series models is based on the Apache 2.0 agreement, while the model weight is based on the BAAI Aquila Model License Agreement. Users can use it for commercial purposes as long as they meet the licensing restrictions.
✍️ Possesses Chinese and English knowledge. The Aquila series model is trained from scratch on a high-quality corpus of Chinese and English languages, with Chinese corpora accounting for about 40%, ensuring that the model accumulates native Chinese world knowledge during the pre-training phase, rather than translated knowledge.
👮♀️ Complies with domestic data regulations. The Chinese corpora of the Aquila series models come from Intelligence Source's accumulated Chinese datasets over the years, including Chinese internet data from over 10,000 sources (more than 99% of which are domestic sources), as well as high-quality Chinese literature and book data supported by authoritative domestic organizations. We will continue to accumulate high-quality and diverse datasets and incorporate them into the subsequent training of the Aquila base models.
🎯 Continuous improvements and open sourcing. We will continue to improve training data, optimize training methods, and enhance model performance, cultivate a flourishing "model tree" on a better base model foundation, and continuously update open-source versions.
The additional details of the Aquila model will be presented in the official technical report. Please stay tuned for updates on official channels, including the FlagAI GitHub repository, FlagAI's Zhihu account and FlagAI's official technical communication group.
| Model | Model Type | Description | Status | GPUs Used |
|---|---|---|---|---|
| Aquila-7B | Base model, 7 billion parameters | Aquila Base Model inherits the architectural design advantages of GPT-3 and LLaMA. It replaces a batch of more efficient underlying operator implementations, redesigns the implementation of bilingual tokenizer, upgrades BMTrain parallel training method, and achieves nearly 8 times the training efficiency of Magtron+DeepSpeed ZeRO-2. | Released | Nvidia-A100 |
| Aquila-33B | Base model, 33 billion parameters | Same as above | Coming soon | Nvidia-A100 |
| AquilaChat-7B | SFT model, fine-tuned and RL based on Aquila-7B | AquilaChat Dialog Model supports fluent text dialogue and multiple language generation tasks, and realizes the call of AquilaChat to other models and tools by defining an expandable special instruction specification, which is easy to extend. For example, calling the open source AltDiffusion multimodal language image generation model of Flagship Intelligence achieved smooth image generation capability. Together with Flagship Intelligence's InstructFace multi-step controllable text-picture model, it is easy to achieve multi-step controllable editing of human face images. | Released | Nvidia-A100 |
| AquilaChat-33B | SFT model, fine-tuned and RL based on Aquila-33B | Same as above | Coming soon | Nvidia-A100 |
| AquilaCode-7B-NV | Base model, "text-code" generation model, further pre-trained based on Aquila-7B, trained on Nvidia | AquilaCode-7B achieves high performance with small data sets and parameters, and is currently the best open source code model that supports both Chinese and English, trained using training code data with compliant open source licenses after high-quality filtering. AquilaCode-7B has been trained on both Nvidia and domestic chips for code models. | Released | Nvidia-A100 |
| AquilaCode-7B-TS | Base model, "text-code" generation model, further pre-trained based on Aquila-7B, trained on Horizon Robotics chips | Same as above | Released | Tianshu-BI-V100 |
We will continue to release improved versions of Aquila model as open source. For more details, please refer to the Change Log.
Quick Start AquilaChat-7B（Chat model）
1. Inference
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
from cyg_conversation import covert_prompt_to_input_ids_with_history
tokenizer = AutoTokenizer.from_pretrained("BAAI/AquilaChat-7B")
model = AutoModelForCausalLM.from_pretrained("BAAI/AquilaChat-7B")
model.eval()
model.to("cuda:0")
vocab = tokenizer.vocab
print(len(vocab))
text = "请给出10个要到北京旅游的理由。"
tokens = covert_prompt_to_input_ids_with_history(text, history=[], tokenizer=tokenizer, max_token=512)
tokens = torch.tensor(tokens)[None,].to("cuda:0")
with torch.no_grad():
out = model.generate(tokens, do_sample=True, max_length=512, eos_token_id=100007)[0]
out = tokenizer.decode(out.cpu().numpy().tolist())
print(out)
usning NBCE Inference
import json
import torch
from transformers import AutoTokenizer
from transformers import AutoModelForCausalLM
from transformers import TopPLogitsWarper, LogitsProcessorList
import pdb
# load tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_path)
tokenizer.padding_side = 'left'
tokenizer.pad_token = tokenizer.unk_token
# load Aquila model
model = AutoModelForCausalLM.from_pretrained(model_path, torch_dtype=torch.float16)
device = torch.device('cuda')
model.to(device)
# load example Context
from cyg_conversation import default_conversation
conv = default_conversation.copy()
contexts = json.load(open('code_text_2.json'))
question = "请解释这段程序的功能："
batch = []
conv.append_message(conv.roles[0], question)
conv.append_message(conv.roles[1], None)
batch.append(conv.get_prompt())
# concat context and question
for ci,context in enumerate(contexts):
conv1 = default_conversation.copy()
conv1.append_message(conv.roles[0], context+question)
conv1.append_message(conv.roles[1], None)
batch.append(conv1.get_prompt())
print('Context长度分布：', [len(text) for text in batch])
print('Context总长度：', sum([len(text) for text in batch]))
# Top-P
processors = LogitsProcessorList()
processors.append(TopPLogitsWarper(0.95))
# Copied from https://github.com/bojone/NBCE/blob/main/test.py#L51-L106
@torch.inference_mode()
def generate(max_tokens):
"""Naive Bayes-based Context Extension example code
"""
inputs = tokenizer(batch, padding='longest', return_tensors='pt').to(device)
input_ids = inputs.input_ids
attention_mask = inputs.attention_mask
print('input_ids', input_ids.shape)
past_key_values = None
n = input_ids.shape[0]
for i in range(max_tokens):
# model output
outputs = model(input_ids=input_ids,
attention_mask=attention_mask,
return_dict=True,
use_cache=True,
past_key_values=past_key_values
)
past_key_values = outputs.past_key_values
# ===== NBCE core code starts =====
beta, eta = 0.25, 0.1
logits = outputs.logits[:, -1]
logits = logits - logits.logsumexp(dim=-1, keepdims=True)
logits = processors(input_ids, logits)
entropy = -(logits.exp() * logits.clip(-100, 0)).sum(dim=-1)
if i > 0:
entropy[k] -= eta
k = entropy[1:].argmin() + 1
logits_max = logits[k]
logits_uncond = logits[0]
logits_merged = (1 + beta) * logits_max - beta * logits_uncond
logits = torch.where(logits_uncond > -100, logits_merged, logits_max)
# ===== NBCE core code ends =====
# Building a distribution and sampling
# tau = 1 is standard random sampling，tau->0 is greedy search
# For simplicity, top-k and top-p truncation are not implemented here.
tau = 0.01
probas = torch.nn.functional.softmax(logits[None] / tau , dim=-1)
next_tokens = torch.multinomial(probas, num_samples=1).squeeze(1)
if next_tokens[0] == tokenizer.eos_token_id:
break
ret = tokenizer.batch_decode(next_tokens)
print(ret[0], flush=True, end='')
# prepare for next iteration
input_ids = next_tokens.unsqueeze(-1).tile(n, 1)
attention_mask = torch.cat([attention_mask, torch.ones(n, 1, dtype=torch.long, device=device)], dim=-1)
if __name__ == '__main__':
generate(1000)
License
AquilaChat-7B and AquilaChat-33B open-source model is licensed under BAAI Aquila Model Licence Agreement