ericrisco commited on
Commit
b3d7e79
·
verified ·
1 Parent(s): a97f4e4

Dataset card for 20261104-0036-12b-k-canon

Browse files
Files changed (1) hide show
  1. README.md +12 -3
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 exact data of training run **`20261102-1437-e4b-k-canon`** (variant `K-canon`).
 
 
 
28
 
29
  | Split | Conversations |
30
  |---|---:|
31
- | train | 28679 |
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
- Questions of the maia-bench benchmark and the articles it reserves are excluded.
 
 
 
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