Second multi-task round: hard negatives for the description question, breaking API merged into breaking

README.md

@@ -127,6 +127,13 @@ The test split holds about 10% of the rows of each config, chosen by **whole rep
 
 There is no split: train and test rows are together, and you split by `grp`. `multitask/goal.txt` is the goal we give the typestrict.ai structuring agent for this file.
 
+`multitask/ts-prguard-v2.jsonl.gz` (41,606 rows, 6 tasks) is the second round, with two changes:
+
+- `pr-descricao-noul` also has hard negatives: for 2,706 real pairs, the same diff with a plausible description that misstates it (another change to the same code, only a minor part of it, the opposite effect, or a refactoring that changes behavior). The descriptions were written by a language model from the public commit; the label comes from how they were built. In our smoke run they took the model from 0.61 to 0.91 AUC on such cases while the easy negatives stayed within noise;
+- `prg-quebra-api` is merged into `prg-quebra` (same question and options).
+
+`multitask/goal-v2.txt` is its goal.
+
 ## 🗂️ Fields
 
 <div class="dark:hidden" style="overflow-x:auto;">

multitask/goal-v2.txt

@@ -0,0 +1 @@
+The file is already in training format: do not change, trim, rewrite or join any value. Make exactly one task per distinct value of the task column, named as that value, with that task's kind and question columns as they are. kind noul: fixed labels with empty descriptions, in this order: options[1] then options[2] (the no, then the yes), and SQL: select state, label, grp, lang from <table> where task = '<task>'. kind choice: no fixed labels (options per example), SQL: select state, options, answer, grp, lang from <table> where task = '<task>'. Tasks: pr-descricao-noul, prg-complexidade, prg-quebra, prg-testes, prg-tipo, prg-vul.

multitask/ts-prguard-v2.jsonl.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a5dbed6495c6625d1180208cbdfd7b740cc34b0ea37e87f3653d215668aeeaa
+size 40552387