rankalign-v6-gemma-2-9b-it-d0.15-e2-ambigqa-all-tcs-fsx-lo0.1

Fine-tuned checkpoint from the rankalign project.

Training Details

Field	Value
Base model	google/gemma-2-9b-it
Version	v6
Task	ambigqa-all
Epoch	2
Delta	0.15
Typicality correction	self
Length normalization	False
Preference loss weight	1
NLL validator weight	0
NLL generator weight	0
Validator log-odds	False
Force same-x	True
Semi-supervised ratio	None
Labeled-only ratio	0.1

Reproducibility

Original checkpoint name: v6-google--gemma-2-9b-it-delta0.15-epoch2--ambigqa-all--d2g--random--alpha1.0--tc-self--full-completion--force-same-x--labelonly0.1

To evaluate:

python scripts/eval_by_claude.py \
    --model TAUR-dev/rankalign-v6-gemma-2-9b-it-d0.15-e2-ambigqa-all-tcs-fsx-lo0.1 \
    --task ambigqa-all \
    --split_type random --gen-shots zero --disc-shots few --validator-log-odds --save-scores-csv \
    --self-typicality