Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -14,16 +14,40 @@ fallacies of Zanartu et al. 2024 (doi.org/10.1038/s41598-024-76139-w).
|
|
| 14 |
|
| 15 |
| path | rows | assistant turn |
|
| 16 |
|---|---:|---|
|
| 17 |
-
| `recot/train.jsonl` |
|
| 18 |
-
| `recot/train_eval.jsonl` |
|
| 19 |
-
| `labels/train.jsonl` |
|
| 20 |
-
| `labels/train_eval.jsonl` |
|
| 21 |
|
| 22 |
Both arms come from one teacher pass over the same rows with the same stratified
|
| 23 |
split (seed 42), so they differ only in whether the reasoning is present. That
|
| 24 |
is the ablation for the source paper's own thesis: is the manual deconstruction
|
| 25 |
what makes the dataset valuable, or do the labels alone suffice?
|
| 26 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 27 |
## Provenance
|
| 28 |
|
| 29 |
Source rows are the FLICC training split, vendored verbatim from
|
|
|
|
| 14 |
|
| 15 |
| path | rows | assistant turn |
|
| 16 |
|---|---:|---|
|
| 17 |
+
| `recot/train.jsonl` | 1614 | full deconstruction trace + YAML |
|
| 18 |
+
| `recot/train_eval.jsonl` | 177 | full deconstruction trace + YAML |
|
| 19 |
+
| `labels/train.jsonl` | 1614 | YAML answer only |
|
| 20 |
+
| `labels/train_eval.jsonl` | 177 | YAML answer only |
|
| 21 |
|
| 22 |
Both arms come from one teacher pass over the same rows with the same stratified
|
| 23 |
split (seed 42), so they differ only in whether the reasoning is present. That
|
| 24 |
is the ablation for the source paper's own thesis: is the manual deconstruction
|
| 25 |
what makes the dataset valuable, or do the labels alone suffice?
|
| 26 |
|
| 27 |
+
`split_manifest.json` in each folder records the seed, the rule, and which
|
| 28 |
+
itemId went to which split, so the split is checkable without these files.
|
| 29 |
+
|
| 30 |
+
## Decontamination
|
| 31 |
+
|
| 32 |
+
5 of the 1,796 training rows are excluded here. Normalising away punctuation and
|
| 33 |
+
case, then confirming by character diff, three training rows are >=98% identical
|
| 34 |
+
to an evaluation row:
|
| 35 |
+
|
| 36 |
+
| training row | eval row | difference |
|
| 37 |
+
|---|---|---|
|
| 38 |
+
| `train-0605` | `test-0079` | one capital letter, a trailing period |
|
| 39 |
+
| `train-1390` | `val-0382` | curly quotes vs a period |
|
| 40 |
+
| `train-1599` | `val-0113` | one apostrophe |
|
| 41 |
+
|
| 42 |
+
Two further rows, `train-0223` and `train-0611`, are the same sentence carrying
|
| 43 |
+
different gold labels (`single cause` and `misrepresentation`), so the teacher
|
| 44 |
+
produced contradictory traces for one input.
|
| 45 |
+
|
| 46 |
+
The same conflict exists inside the validation split (`val-0108` /`val-0386`,
|
| 47 |
+
`false equivalence` vs `single cause`) and is left untouched — correcting
|
| 48 |
+
someone else's evaluation annotation is not appropriate; it is reported as label
|
| 49 |
+
noise instead.
|
| 50 |
+
|
| 51 |
## Provenance
|
| 52 |
|
| 53 |
Source rows are the FLICC training split, vendored verbatim from
|