Update README.md
Browse files
README.md
CHANGED
@@ -16,4 +16,33 @@ task_categories:
|
|
16 |
- text-generation
|
17 |
- text-classification
|
18 |
- question-answering
|
19 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
16 |
- text-generation
|
17 |
- text-classification
|
18 |
- question-answering
|
19 |
+
---
|
20 |
+
|
21 |
+
# ScienceWikiSmallChunk
|
22 |
+
|
23 |
+
Processed version of millawell/wikipedia_field_of_science, prepared to be used in small context length RAG systems. Chunk length is tokenizer dependent, but each chunk should be around 512 tokens. Longer wikipedia pages have been split into smaller entries, with title added as a prefix.
|
24 |
+
|
25 |
+
If you wish to prepare some other chunk length:
|
26 |
+
- use millawell/wikipedia_field_of_science
|
27 |
+
- adapt chunker function:
|
28 |
+
```
|
29 |
+
def chunker_clean(results, example, length=512, approx_token=3, prefix=""):
|
30 |
+
if len(results) == 0:
|
31 |
+
regex_pattern = r'[\n\s]*\n[\n\s]*'
|
32 |
+
example = re.sub(regex_pattern, " ", example).strip().replace(prefix, "")
|
33 |
+
chunk_length = length * approx_token
|
34 |
+
if len(example) > chunk_length:
|
35 |
+
first = example[:chunk_length]
|
36 |
+
chunk = ".".join(first.split(".")[:-1])
|
37 |
+
if len(chunk) == 0:
|
38 |
+
chunk = first
|
39 |
+
rest = example[len(chunk)+1:]
|
40 |
+
results.append(prefix+chunk.strip())
|
41 |
+
if len(rest) > chunk_length:
|
42 |
+
chunker_clean(results, rest.strip(), length=length, approx_token=approx_token, prefix=prefix)
|
43 |
+
else:
|
44 |
+
results.append(prefix+rest.strip())
|
45 |
+
else:
|
46 |
+
results.append(prefix+example.strip())
|
47 |
+
return results
|
48 |
+
```
|