Skip to content

Handle outputs segmented into multiple units (with ProbScalarizedModel only)#11

Merged
dennislwei merged 22 commits intomainfrom
mexgen
Jan 6, 2026
Merged

Handle outputs segmented into multiple units (with ProbScalarizedModel only)#11
dennislwei merged 22 commits intomainfrom
mexgen

Conversation

@dennislwei
Copy link
Member

@dennislwei dennislwei commented Jan 5, 2026

  1. Segment output text or take pre-segmented output
  2. Attribute each unit of output to input units
  3. Evaluate attribution of each output unit using PerturbCurveEvaluator

NOTE: Items 2 and 3 are implemented only for the ProbScalarizedModel scalarizer.

…not a batch)

Signed-off-by: Dennis Wei <dwei@us.ibm.com>
…larizedModel

- Find token boundaries of output units
- Aggregate log probs within each output unit

Signed-off-by: Dennis Wei <dwei@us.ibm.com>
- GeneratedOutput object should always contain output_text in addition to output_ids

Signed-off-by: Dennis Wei <dwei@us.ibm.com>
…larizedModel only

Signed-off-by: Dennis Wei <dwei@us.ibm.com>
Signed-off-by: Dennis Wei <dwei@us.ibm.com>
Signed-off-by: Dennis Wei <dwei@us.ibm.com>
explainer

Signed-off-by: Dennis Wei <dwei@us.ibm.com>
…ith ProbScalarizedModel only

Signed-off-by: Dennis Wei <dwei@us.ibm.com>
Signed-off-by: Dennis Wei <dwei@us.ibm.com>
Signed-off-by: Dennis Wei <dwei@us.ibm.com>
- look for token in next unit as well in order to skip split tokens and special tokens

Signed-off-by: Dennis Wei <dwei@us.ibm.com>
Signed-off-by: Dennis Wei <dwei@us.ibm.com>
- Remove search for current token in next unit, only check whether current unit is long enough

Signed-off-by: Dennis Wei <dwei@us.ibm.com>
This reverts commit ae36906.

Signed-off-by: Dennis Wei <dwei@us.ibm.com>
Signed-off-by: Dennis Wei <dwei@us.ibm.com>
Signed-off-by: Dennis Wei <dwei@us.ibm.com>
Signed-off-by: Dennis Wei <dwei@us.ibm.com>
Signed-off-by: Dennis Wei <dwei@us.ibm.com>
Signed-off-by: Dennis Wei <dwei@us.ibm.com>
Signed-off-by: Dennis Wei <dwei@us.ibm.com>
Signed-off-by: Dennis Wei <dwei@us.ibm.com>
@dennislwei dennislwei merged commit 5bd1c76 into main Jan 6, 2026
2 checks passed
dennislwei added a commit to emiehling/ICX360 that referenced this pull request Jan 8, 2026
Signed-off-by: Dennis Wei <dwei@us.ibm.com>
dennislwei added a commit to emiehling/ICX360 that referenced this pull request Jan 8, 2026
…compatible with PR IBM#11

Signed-off-by: Dennis Wei <dwei@us.ibm.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant