Refactor evaluation to use Inspect, bump package versions, rename package to steerability - #25

Merged

emiehling merged 33 commits into

generative-computing:mainfrom

emiehling:eval-refactor

Sep 8, 2026

Commits on Aug 22, 2026

Port to transformers v5 and bump the ML dependency stack
emiehling
committed
94dcc71

Commits on Aug 24, 2026

Replace the evaluation stack with Inspect AI
emiehling
committed
8725041

Commits on Aug 26, 2026

Add the .spipe serialization format and broaden model-family support
emiehling
committed
baabdc5

Commits on Aug 27, 2026

Commits on Aug 28, 2026

Port the studies to the Inspect stack and add PASTA head profiling
emiehling
committed
c059457

Commits on Sep 1, 2026

Add system-prompt and user-prefix controls and reorganize the notebooks
emiehling
committed
cca3008

Commits on Sep 2, 2026

Harden the vLLM boot and serve environments and the TRL wrappers
emiehling
committed
0cdd960

Commits on Sep 3, 2026

Commits on Sep 4, 2026

Canonicalize the elements of NumPy object arrays
emiehling
committed
d53be01
Merge remote-tracking branch 'upstream/main' into eval-refactor
emiehling
committed
7c96c89
Freeze every control before encoding recipe args in build_spipe
emiehling
committed
70b3ebc
Let the manifest record govern the verification wrap of a decoded artifact
emiehling
committed
23edce5
Honour artifacts= when saving onto a bundle's own directory
emiehling
committed
f2b59ee
Classify artifact availability per artifact in SPipe.verify
emiehling
committed
201f611
Sort $map entries in the spipe codec
emiehling
committed
f01c4a7
Accept a close-only continuation in the token-mode reasoning split
emiehling
committed
4e093e6
Raise whenever eval_set reports failure in InspectSuite.run
emiehling
committed
90c0fe3
Check the output count of a batched dispatch in the collator
emiehling
committed
d4ae748
Document the score order of ProbeEvaluation
emiehling
committed
23d513c
Correct the _config_layout docstring on missing structural facts
emiehling
committed
b73ed98
Warn when trial seeds cannot be attached for lack of a temperature
emiehling
committed
65408ea
Read the reward head at the candidate position in RewardModelValue
emiehling
committed
332b245
Compute the prefix reward loss in float32
emiehling
committed
daa1dbf
Refuse a padded prefix mask in CandidateForward
emiehling
committed
241ea7a
Run the pad-id candidate test on the Llama classifier
emiehling
committed
11a1352
Classify GenerateConfig.stream_idle_timeout as an upstream field
emiehling
committed
95521ec
Refactor granite sequence classifier resolution
emiehling
committed
f54a385

Commits on Sep 5, 2026

Revise the dependency groups in pyproject.toml
emiehling
committed
285faf8

Commits on Sep 8, 2026

Log the constructor base_seed in the unset-temperature branch
emiehling
committed
83f80bd