yueyulin/unit_extraction_states

This checkpoint is a states tuning file from RWKV-6-7B. Please download the base model from https://huggingface.co/BlinkDL/rwkv-6-world/tree/main . It will extract number and unit from input Chinese text. Usage:

update the latest rwkv package: pip install --upgrade rwkv

Download the base model and the states file. You may download the states from the epoch_2 directory.

Following the codes:
    Loading the model and states

from rwkv.model import RWKV
from rwkv.utils import PIPELINE, PIPELINE_ARGS
import torch

# download models: https://huggingface.co/BlinkDL
model = RWKV(model='/media/yueyulin/KINGSTON/models/rwkv6/RWKV-x060-World-7B-v2.1-20240507-ctx4096.pth', strategy='cuda fp16')
print(model.args)
pipeline = PIPELINE(model, "rwkv_vocab_v20230424") # 20B_tokenizer.json is in https://github.com/BlinkDL/ChatRWKV
# use pipeline = PIPELINE(model, "rwkv_vocab_v20230424") for rwkv "world" models
states_file = '/media/yueyulin/data_4t/models/states_tuning/custom_trainer/epoch_2/RWKV-x060-World-7B-v2.1-20240507-ctx4096.pth.pth'
states = torch.load(states_file)
states_value = []
device = 'cuda'
n_head = model.args.n_head
head_size = model.args.n_embd//model.args.n_head
for i in range(model.args.n_layer):
    key = f'blocks.{i}.att.time_state'
    value = states[key]
    prev_x = torch.zeros(model.args.n_embd,device=device,dtype=torch.float16)
    prev_states = value.clone().detach().to(device=device,dtype=torch.float16).transpose(1,2)
    prev_ffn = torch.zeros(model.args.n_embd,device=device,dtype=torch.float16)
    states_value.append(prev_x)
    states_value.append(prev_states)
    states_value.append(prev_ffn)

• Try the following examples:

cat_char = '🐱'
bot_char = '🤖'
instruction ='你是一个单位提取专家。请从input中抽取出数字和单位，请按照JSON字符串的格式回答，无法提取则不输出。'
input_text = '大约503万平方米'
ctx = f'{cat_char}:{instruction}\n{input_text}\n{bot_char}:'
print(ctx)

def my_print(s):
    print(s, end='', flush=True)

# For alpha_frequency and alpha_presence, see "Frequency and presence penalties":
# https://platform.openai.com/docs/api-reference/parameter-details

args = PIPELINE_ARGS(temperature = 1.0, top_p = 0, top_k = 0, # top_k = 0 then ignore
                     alpha_frequency = 0.25,
                     alpha_presence = 0.25,
                     alpha_decay = 0.996, # gradually decay the penalty
                     token_ban = [0], # ban the generation of some tokens
                     token_stop = [0,1], # stop generation whenever you see any token here
                     chunk_len = 256) # split input into chunks to save VRAM (shorter -> slower)

pipeline.generate(ctx, token_count=200, args=args, callback=my_print,state=states_value)
print('\n')
input_text = '4845人'
...
input_text = '约89434户'
...
input_text = '可能有38.87亿平方公里'
...

The output should look like:

{"number": 5030000, "unit": "平方米"}
{"number": 4845, "unit": "人"}
{"number": 89434, "unit": "户"}
{"number": 3887000000, "unit": "平方公里"}