Coding-With-Bashir commited on
Commit
b7181ad
·
verified ·
1 Parent(s): 6039777

Upload .\docs\data-collection.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. .//docs//data-collection.md +39 -0
.//docs//data-collection.md ADDED
@@ -0,0 +1,39 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Data Collection
2
+
3
+ ## Overview
4
+
5
+ BwengeAi collects Kinyarwanda language data from 200+ sources across 12 pipeline stages.
6
+
7
+ ## Pipeline Stages
8
+
9
+ | Step | Collector | Source | Description |
10
+ |------|-----------|--------|-------------|
11
+ | 1 | HuggingfaceCollector | HuggingFace Hub | 200 datasets via streaming |
12
+ | 2 | ParallelCorpusCollector | HuggingFace Hub | michsethowusu parallel pairs |
13
+ | 3 | CC100OSCARCollector | HuggingFace Hub | CC-100/OSCAR monolingual |
14
+ | 4 | MasakhaneCollector | HuggingFace Hub | African NLP benchmarks |
15
+ | 5 | WikipediaCollector | rw.wikipedia.org | API + XML dump |
16
+ | 6 | IgiheScraper | igihe.com | News articles |
17
+ | 7 | KigaliTodayScraper | kigalitoday.com | News articles |
18
+ | 8 | RSSCollector | newtimes.co.rw | RSS feed articles |
19
+ | 9 | WikisourceCollector | rw.wikisource.org | Public texts |
20
+ | 10 | WiktionaryCollector | rw.wiktionary.org | Dictionary entries |
21
+ | 11 | RBCCollector | rbc.gov.rw | Health content |
22
+ | 12 | DataProcessor | Local files | Clean + deduplicate |
23
+
24
+ ## Running
25
+
26
+ ```bash
27
+ python scripts/collect_all.py
28
+ ```
29
+
30
+ ## Output
31
+
32
+ - `data/raw/` — Raw collected JSONL files
33
+ - `data/processed/training_data.jsonl` — Cleaned text for LM training
34
+ - `data/processed/instruction_data.jsonl` — Instruction-following format
35
+ - `data/processed/chat_data.jsonl` — Chat/conversation format
36
+
37
+ ## Configuration
38
+
39
+ All dataset sources are defined in `configs/default.yaml`.