Datasets:
Recover cached SimRAG paraphrases with minimal selection
Browse files- README.md +3 -3
- retrieved_answer_judged/multi_source.jsonl +0 -0
- retrieved_answer_judged/multiple_choice.jsonl +0 -0
- retrieved_answer_judged/short_span.jsonl +0 -0
- retrieved_answer_judged/unanswerable.jsonl +0 -0
- retrieved_selection/multi_source.jsonl +0 -0
- retrieved_selection/multiple_choice.jsonl +0 -0
- retrieved_selection/short_span.jsonl +0 -0
- retrieved_selection/unanswerable.jsonl +0 -0
README.md
CHANGED
|
@@ -29,15 +29,15 @@ configs:
|
|
| 29 |
path: retrieved_answer_judged/multi_source.jsonl
|
| 30 |
---
|
| 31 |
|
| 32 |
-
# SimRAG Tahap II —
|
| 33 |
|
| 34 |
QA dan konteks retrieval berasal dari run B. Keanggotaan dua dataset dipertahankan.
|
| 35 |
-
Seleksi parafrase: `
|
| 36 |
Tidak memaksakan anti-salin, metadata, overlap kata, letak span di kalimat pertama, atau penghapusan kalimat pendukung.
|
| 37 |
`response` adalah jawaban awal. `response_paraphrase` adalah target jawaban hasil pemulihan.
|
| 38 |
Untuk training gunakan `prompt_paraphrase` (MS: soal gabungan alami; jenis lain: sama dengan prompt),
|
| 39 |
`response_paraphrase`, `specific_instruction`, dan `context`.
|
| 40 |
-
Seleksi
|
| 41 |
Audit pemulihan disimpan lokal agar tidak menjadi split dataset tambahan.
|
| 42 |
|
| 43 |
Jumlah per config:
|
|
|
|
| 29 |
path: retrieved_answer_judged/multi_source.jsonl
|
| 30 |
---
|
| 31 |
|
| 32 |
+
# SimRAG Tahap II — relaxed paraphrase recovery
|
| 33 |
|
| 34 |
QA dan konteks retrieval berasal dari run B. Keanggotaan dua dataset dipertahankan.
|
| 35 |
+
Seleksi parafrase: `minimal`. Keluaran mentah dipulihkan memakai hash prompt yang sama.
|
| 36 |
Tidak memaksakan anti-salin, metadata, overlap kata, letak span di kalimat pertama, atau penghapusan kalimat pendukung.
|
| 37 |
`response` adalah jawaban awal. `response_paraphrase` adalah target jawaban hasil pemulihan.
|
| 38 |
Untuk training gunakan `prompt_paraphrase` (MS: soal gabungan alami; jenis lain: sama dengan prompt),
|
| 39 |
`response_paraphrase`, `specific_instruction`, dan `context`.
|
| 40 |
+
Seleksi minimal bukan validasi semantik semua fakta. Record tanpa kandidat yang lolos tetap memakai jawaban sebelumnya.
|
| 41 |
Audit pemulihan disimpan lokal agar tidak menjadi split dataset tambahan.
|
| 42 |
|
| 43 |
Jumlah per config:
|
retrieved_answer_judged/multi_source.jsonl
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|
retrieved_answer_judged/multiple_choice.jsonl
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|
retrieved_answer_judged/short_span.jsonl
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|
retrieved_answer_judged/unanswerable.jsonl
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|
retrieved_selection/multi_source.jsonl
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|
retrieved_selection/multiple_choice.jsonl
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|
retrieved_selection/short_span.jsonl
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|
retrieved_selection/unanswerable.jsonl
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|