riccardomusmeci commited on
Commit
348a989
1 Parent(s): 370c79d

Update README

Browse files
Files changed (1) hide show
  1. README.md +105 -0
README.md CHANGED
@@ -1,3 +1,108 @@
1
  ---
2
  license: mit
 
 
 
 
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: mit
3
+ language:
4
+ - en
5
+ pipeline_tag: feature-extraction
6
+ tags:
7
+ - e5-mistral-7b-instruct
8
+ - mlx-llm
9
+ - mlx
10
+ - feature-extraction
11
+ - embeddings
12
+ library_name: mlx-llm
13
  ---
14
+
15
+
16
+ # E5-mistral-7b-instruct
17
+
18
+ [Improving Text Embeddings with Large Language Models](https://arxiv.org/pdf/2401.00368.pdf). Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, Furu Wei, arXiv 2024
19
+
20
+ This model has 32 layers and the embedding size is 4096.
21
+
22
+ ## Model description
23
+
24
+ Please, refer to the [original model card](https://huggingface.co/intfloat/e5-mistral-7b-instruct) for more details on E5-mistral-7b-instruct.
25
+
26
+ ## Use with mlx-llm
27
+
28
+ Download weights from files section and install `mlx-llm` from GitHub.
29
+ ```bash
30
+ git clone https://github.com/riccardomusmeci/mlx-llm
31
+ cd mlx-llm
32
+ pip install .
33
+ ```
34
+
35
+ Run
36
+
37
+ ```python
38
+ import mlx.core as mx
39
+ import numpy as np
40
+ from mlx_llm.model import create_model
41
+ from transformers import AutoTokenizer
42
+
43
+ model = create_model(
44
+ "e5-mistral-7b-instruct",
45
+ weights_path="path/to/weights.npz",
46
+ strict=False
47
+ )
48
+
49
+ def get_detailed_instruct(task_description: str, query: str) -> str:
50
+ return f'Instruct: {task_description}\nQuery: {query}'
51
+
52
+ def last_token_pool(embeds: mx.array, attn_mask: mx.array) -> mx.array:
53
+ left_padding = (attn_mask[:, -1].sum() == attn_mask.shape[0])
54
+ if left_padding:
55
+ return embeds[:, -1]
56
+ else:
57
+ sequence_lengths = attn_mask.sum(axis=1) - 1
58
+ batch_size = embeds.shape[0]
59
+ return embeds[mx.arange(batch_size), sequence_lengths]
60
+
61
+ task = 'Given a web search query, retrieve relevant passages that answer the query'
62
+
63
+ input_texts = [
64
+ get_detailed_instruct(task, 'how much protein should a female eat'),
65
+ # get_detailed_instruct(task, 'summit define'),
66
+ "As a general guideline, the CDC's average requirement of protein for women ages 19 to 70 is 46 grams per day. But, as you can see from this chart, you'll need to increase that if you're expecting or training for a marathon. Check out the chart below to see how much protein you should be eating each day.",
67
+ "Definition of summit for English Language Learners. : 1 the highest point of a mountain : the top of a mountain. : 2 the highest level. : 3 a meeting or series of meetings between the leaders of two or more governments."
68
+ ]
69
+
70
+ tokenizer = AutoTokenizer.from_pretrained('intfloat/e5-mistral-7b-instruct')
71
+
72
+ # prepare input and attn_mask
73
+ max_length = 4096
74
+ batch_dict = tokenizer(
75
+ input_texts,
76
+ max_length=max_length - 1,
77
+ return_attention_mask=False,
78
+ padding=False,
79
+ truncation=True
80
+ )
81
+
82
+ batch_dict['input_ids'] = [
83
+ input_ids + [tokenizer.eos_token_id] for input_ids in batch_dict['input_ids']
84
+ ]
85
+
86
+ batch_dict = tokenizer.pad(
87
+ batch_dict,
88
+ padding=True,
89
+ return_attention_mask=True,
90
+ return_tensors='np'
91
+ )
92
+
93
+ x = mx.array(batch_dict["input_ids"].tolist())
94
+ attn_mask = mx.array(batch_dict["attention_mask"].tolist())
95
+
96
+ # compute embed
97
+ embeds = model.embed(x)
98
+ mx.eval(embeds)
99
+
100
+ embeds = np.array(last_token_pool(embeds, attn_mask))
101
+
102
+ # Normalize embeds
103
+ norm_den = np.linalg.norm(embeds, axis=-1)
104
+ norm_embeds = embeds / norm_den[:, None]
105
+
106
+ scores = (norm_embeds @ norm_embeds.T) * 100
107
+ print(scores)
108
+ ```