File size: 2,474 Bytes
8241242
 
 
 
 
626e1a2
8241242
 
 
 
 
 
 
 
2ce7d81
626e1a2
2ce7d81
8241242
 
 
 
 
 
b10697f
626e1a2
8241242
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
---
base_model: nisten/Biggie-SmoLlm-0.15B-Base
license: mit
datasets:
- LDJnr/Capybara
- andthattoo/subqueries
pipeline_tag: text-generation
tags:
- llama
---


### Fine-tuned [Biggie-SmoLlm-0.15B-Base](https://huggingface.co/nisten/Biggie-SmoLlm-0.15B-Base) for generating subqueries

This dude is trained for boosting the performance of your RAG based question answering app

My motivation was to tackle a core problem of RAG with an extremely lightweight, but capable model.

If queries are 
- multi-hop logic, break into simpler subqueries that focuses on a different step
- vague, ask follow up questions
- multiple sub questions, generate multiple queries for each of them

Training data was generated with [Dria](https://dria.co): A decentralized p2p network for synthetic data. Join [discord](https://discord.gg/dria) to help decentralized data generation.


Heads up: [Ollama](https://ollama.com/andthattoo/subquery-smollm) version works 160 tps on 1 CPU core. No GPU? No worries. This little dude’s got you. 

Use the model:

```python
from transformers import AutoModel, AutoConfig, AutoTokenizer, AutoModelForCausalLM

config = AutoConfig.from_pretrained("andthattoo/subquery-SmolLM")
tokenizer = AutoTokenizer.from_pretrained("andthattoo/subquery-SmolLM")
model = AutoModelForCausalLM.from_pretrained("andthattoo/subquery-SmolLM", torch_dtype=torch.bfloat16)

if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token
    model.config.pad_token_id = model.config.eos_token_id

input_data = "Generate subqueries for a given question. <question>What is this?</question>"
inputs = tokenizer(input_data, return_tensors='pt')
output = model.generate(**inputs, max_new_tokens=100)
decoded_output = tokenizer.decode(output[0], skip_special_tokens=True)
```

Also created a python package for ease of use
```python
pip install subquery
```

```python
from subquery import TransformersSubqueryGenerator

# Using the Transformers backend
generator = TransformersSubqueryGenerator()
result = generator.generate("What is this?")

print("Follow-up questions:", result.follow_up)
print("Subqueries:", result.subquery)
```

or

```python
from subquery import OllamaSubqueryGenerator
# Using the Ollama backend
generator = OllamaSubqueryGenerator()
result = generator.generate("Are the Indiana Harbor and Ship Canal and the Folsom South Canal in the same state?")

print("Follow-up questions:", result.follow_up)
print("Subqueries:", result.subquery)
```