Datasets:
Dataset card for 20261104-0036-12b-k-canon
Browse files
README.md
CHANGED
|
@@ -24,11 +24,14 @@ configs:
|
|
| 24 |
# maia training data
|
| 25 |
|
| 26 |
Conversations in Andorran Catalan used to fine-tune **maia**, a Gemma 4 model specialised in
|
| 27 |
-
Andorra. This version is the
|
|
|
|
|
|
|
|
|
|
| 28 |
|
| 29 |
| Split | Conversations |
|
| 30 |
|---|---:|
|
| 31 |
-
| train |
|
| 32 |
| validation | 732 |
|
| 33 |
|
| 34 |
Each row is `{"messages": [...]}` in chat format. Every conversation starts with the same
|
|
@@ -43,10 +46,16 @@ system message, which the model also expects at inference time:
|
|
| 43 |
- An **augmented canon of facts**: verified atomic facts, each asked about in about 11 different
|
| 44 |
ways (direct, reverse, paraphrase, context, two-turn, false premise, abstention), plus
|
| 45 |
contrast examples between facts that are easy to confuse.
|
|
|
|
|
|
|
|
|
|
| 46 |
- `probe.jsonl`: 338 held-out questions (never in train) with the keywords a correct answer
|
| 47 |
must contain. They measure fact recall before and after training.
|
| 48 |
|
| 49 |
-
|
|
|
|
|
|
|
|
|
|
| 50 |
|
| 51 |
## Sources and licence
|
| 52 |
|
|
|
|
| 24 |
# maia training data
|
| 25 |
|
| 26 |
Conversations in Andorran Catalan used to fine-tune **maia**, a Gemma 4 model specialised in
|
| 27 |
+
Andorra. This version is the data of training run **`20261104-0036-12b-k-canon`** (variant `K-canon`), except for 43
|
| 28 |
+
multiple-choice conversations that were trained on but are not published: a wrong option
|
| 29 |
+
names a person who could not be confirmed as a public figure (the assistant turn is only a
|
| 30 |
+
letter, and the loss is computed on it alone).
|
| 31 |
|
| 32 |
| Split | Conversations |
|
| 33 |
|---|---:|
|
| 34 |
+
| train | 30245 |
|
| 35 |
| validation | 732 |
|
| 36 |
|
| 37 |
Each row is `{"messages": [...]}` in chat format. Every conversation starts with the same
|
|
|
|
| 46 |
- An **augmented canon of facts**: verified atomic facts, each asked about in about 11 different
|
| 47 |
ways (direct, reverse, paraphrase, context, two-turn, false premise, abstention), plus
|
| 48 |
contrast examples between facts that are easy to confuse.
|
| 49 |
+
- Multiple-choice conversations about the same facts, generated by Gemini, whose assistant turn
|
| 50 |
+
is only the letter of the correct option. The wrong options are invented on purpose, so they
|
| 51 |
+
are not facts. Do not read them as statements about the people or things they name.
|
| 52 |
- `probe.jsonl`: 338 held-out questions (never in train) with the keywords a correct answer
|
| 53 |
must contain. They measure fact recall before and after training.
|
| 54 |
|
| 55 |
+
The articles reserved by the maia-bench benchmark are excluded, and so are its questions and
|
| 56 |
+
near-copies of them (word overlap of 0.5 or more). Practice questions can still resemble a
|
| 57 |
+
benchmark question because both ask about the same facts, and that inflates benchmark scores.
|
| 58 |
+
See the evaluation in the model card.
|
| 59 |
|
| 60 |
## Sources and licence
|
| 61 |
|