Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -22,9 +22,9 @@ persona. Generated from 44 EPUB volumes with an LLM extraction pipeline.
|
|
| 22 |
|
| 23 |
| File | Rows | Notes |
|
| 24 |
|---|---|---|
|
| 25 |
-
| `train.jsonl` |
|
| 26 |
-
| `val.jsonl` |
|
| 27 |
-
| `train_meta.jsonl` / `val_meta.jsonl` |
|
| 28 |
| `manifest.json` | -- | counts, split chapters, params, sha256 |
|
| 29 |
| `stats.json` | -- | length distributions |
|
| 30 |
|
|
@@ -46,9 +46,9 @@ At inference, the recommended system message is:
|
|
| 46 |
|
| 47 |
| Lang | Rows |
|
| 48 |
|---|---|
|
| 49 |
-
| `en` |
|
| 50 |
-
| `id` |
|
| 51 |
-
| `jp` |
|
| 52 |
|
| 53 |
User turns are either verbatim dialogue or third-person scene narration prefixed with
|
| 54 |
`[Scene] `; the ratio is roughly 85% / 5%.
|
|
|
|
| 22 |
|
| 23 |
| File | Rows | Notes |
|
| 24 |
|---|---|---|
|
| 25 |
+
| `train.jsonl` | 24723 | `{"messages":[system,user,assistant]}` |
|
| 26 |
+
| `val.jsonl` | 1796 | held out **by chapter**, no leakage |
|
| 27 |
+
| `train_meta.jsonl` / `val_meta.jsonl` | 24723 / 1796 | sidecar: lang, chapter, kind, prompt |
|
| 28 |
| `manifest.json` | -- | counts, split chapters, params, sha256 |
|
| 29 |
| `stats.json` | -- | length distributions |
|
| 30 |
|
|
|
|
| 46 |
|
| 47 |
| Lang | Rows |
|
| 48 |
|---|---|
|
| 49 |
+
| `en` | 8414 |
|
| 50 |
+
| `id` | 8287 |
|
| 51 |
+
| `jp` | 8022 |
|
| 52 |
|
| 53 |
User turns are either verbatim dialogue or third-person scene narration prefixed with
|
| 54 |
`[Scene] `; the ratio is roughly 85% / 5%.
|