Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,72 @@
|
|
1 |
-
---
|
2 |
-
license: afl-3.0
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: afl-3.0
|
3 |
+
datasets:
|
4 |
+
- issai/kazqad
|
5 |
+
language:
|
6 |
+
- kk
|
7 |
+
library_name: transformers
|
8 |
+
pipeline_tag: question-answering
|
9 |
+
---
|
10 |
+
|
11 |
+
# RoBERTa-Large-KazQAD for Question Answering
|
12 |
+
|
13 |
+
## Model Description
|
14 |
+
nur-dev/roberta-large-kazqad is a fine-tuned version of RoBERTa-Kaz-Large, specifically optimized for the Question Answering (QA) task using the Kazakh Open-Domain Question Answering Dataset (KazQAD). This model is trained to extract precise answers from given contexts in the Kazakh language.
|
15 |
+
|
16 |
+
## Usage
|
17 |
+
The model can be used with the Hugging Face Transformers library:
|
18 |
+
```python
|
19 |
+
from transformers import AutoModelForQuestionAnswering, AutoTokenizer
|
20 |
+
import torch
|
21 |
+
|
22 |
+
# Load the fine-tuned model and tokenizer
|
23 |
+
repo_id = 'nur-dev/roberta-large-kazqad'
|
24 |
+
model = AutoModelForQuestionAnswering.from_pretrained(repo_id)
|
25 |
+
tokenizer = AutoTokenizer.from_pretrained(repo_id)
|
26 |
+
|
27 |
+
# Define the context and question
|
28 |
+
context = """
|
29 |
+
Алматы Қазақстанның ең ірі мегаполисі. Алматы – асқақ Тянь-Шань тауы жотасының көкжасыл бауырайынан,
|
30 |
+
Іле Алатауының бөктерінде, Қазақстан Республикасының оңтүстік-шығысында, Еуразия құрлығының орталығында орналасқан қала.
|
31 |
+
Бұл қаланы «қала-бақ» деп те атайды.
|
32 |
+
"""
|
33 |
+
question = "Алматы қаласы Қазақстанның қай бөлігінде орналасқан?"
|
34 |
+
|
35 |
+
# Tokenize the input
|
36 |
+
inputs = tokenizer.encode_plus(
|
37 |
+
question,
|
38 |
+
context,
|
39 |
+
add_special_tokens=True,
|
40 |
+
return_tensors="pt"
|
41 |
+
)
|
42 |
+
|
43 |
+
input_ids = inputs["input_ids"]
|
44 |
+
attention_mask = inputs["attention_mask"]
|
45 |
+
|
46 |
+
# Perform inference
|
47 |
+
with torch.no_grad():
|
48 |
+
outputs = model(input_ids=input_ids, attention_mask=attention_mask)
|
49 |
+
start_logits = outputs.start_logits
|
50 |
+
end_logits = outputs.end_logits
|
51 |
+
|
52 |
+
# Find the answer's start and end position
|
53 |
+
start_index = torch.argmax(start_logits)
|
54 |
+
end_index = torch.argmax(end_logits)
|
55 |
+
|
56 |
+
# Decode the answer from the context
|
57 |
+
answer = tokenizer.decode(input_ids[0][start_index:end_index + 1])
|
58 |
+
|
59 |
+
print(f"Question: {question}")
|
60 |
+
print(f"Answer: {answer}")
|
61 |
+
```
|
62 |
+
|
63 |
+
## Limitations and Biases
|
64 |
+
• Language Specificity: This model is specifically fine-tuned for the Kazakh language and may not perform well in other languages.
|
65 |
+
• Context Length: The model has limitations with very long contexts, as it is fine-tuned for input lengths up to 512 tokens.
|
66 |
+
• Biases: Like other large pre-trained language models, nur-dev/roberta-large-kazqad may exhibit biases present in its training data. Users should be cautious and critically evaluate the model’s outputs, especially for sensitive applications.
|
67 |
+
|
68 |
+
## Model Authors
|
69 |
+
|
70 |
+
**Name:** Kadyrbek Nurgali
|
71 |
+
- **Email:** [email protected]
|
72 |
+
- **LinkedIn:** [Kadyrbek Nurgali](https://www.linkedin.com/in/nurgali-kadyrbek-504260231/)
|