Update README.md
Browse files
README.md
CHANGED
|
@@ -64,6 +64,7 @@ retriever overall and beats Dense LaBSE on every category (nDCG@10 **0.754 vs 0.
|
|
| 64 |
worse than a plain BM25 baseline — exposing a synonym blind spot in dense retrieval on Kazakh.
|
| 65 |
- **BM25 + stemmer never drops below 0.727** on any slice: the most balanced single system.
|
| 66 |
|
|
|
|
| 67 |
End-to-end RAG (Qwen2.5-7B, n=300) is reported honestly: the stemmer raises retrieval
|
| 68 |
hit@3 (0.737 → 0.803), but the end-to-end accuracy gain is **not** statistically
|
| 69 |
significant (McNemar p=0.63) — the bottleneck is the Kazakh-language generator, not the
|
|
|
|
| 64 |
worse than a plain BM25 baseline — exposing a synonym blind spot in dense retrieval on Kazakh.
|
| 65 |
- **BM25 + stemmer never drops below 0.727** on any slice: the most balanced single system.
|
| 66 |
|
| 67 |
+
|
| 68 |
End-to-end RAG (Qwen2.5-7B, n=300) is reported honestly: the stemmer raises retrieval
|
| 69 |
hit@3 (0.737 → 0.803), but the end-to-end accuracy gain is **not** statistically
|
| 70 |
significant (McNemar p=0.63) — the bottleneck is the Kazakh-language generator, not the
|