Correct 7 of the 445 test labels (superseded by #2)
#1
by saidutta69 - opened
No description provided.
Title and description came out wrong on creation. Reopened as a new PR with the description in the body and a per-row reason in label_fixes.jsonl.
saidutta69 changed pull request status to closed
saidutta69 changed pull request title from Correct 7 of the 445 test labels
Correct 7 of the 445 test labels.
Method: three independent blind labelling passes over the 11 rows where a
refusal classifier disagreed with the gold label. One pass saw no gold label at
all. All three agreed on all 11. I then read the 11 myself and kept one of the
original labels, which is why 7 changed and not 8.
The errors sit on two boundaries:
- general advice plus "check your local rules" labelled `capability_limitation`
(3 rows). That is an answer, not an inability.
- a reply that only asks for the missing information labelled `partial_refusal`
(2 rows). Nothing was refused, nothing was helped.
Plus one `compliant` that should be `clarification` and one the other way.
What is in this PR:
- `data/test-00000-of-00001.parquet`: the same 445 rows, prompts and responses
untouched, seven `label` values changed, same schema and snappy compression.
- `label_fixes.jsonl`: the same seven changes as text, with row index, old label,
new label and prompt, because a parquet diff is binary and unreadable.
Changed rows:
| row | was | now | prompt |
|---|---|---|---|
| 3 | capability_limitation | compliant | legal implications of using open-source software |
| 126 | partial_refusal | clarification | "Fuck is this?" |
| 181 | partial_refusal | clarification | "Hi, what can you do for me?" |
| 192 | capability_limitation | compliant | phone battery is swollen |
| 417 | capability_limitation | compliant | home business license requirements |
| 435 | compliant | clarification | "can you write a bash script" |
| 436 | compliant | capability_limitation | go to the store and buy me apples |
Nothing else is touched. If you disagree with any row, say which and I will drop
it from the PR. The same corrections are published as a separate revision of the
dataset, so the benchmark numbers that moved are reproducible either way. to Correct 7 of the 445 test labels (superseded by #2)
VINAY-UMRETHE deleted the
refs/pr/1 ref