CoolFace
Modelpublic

thoughtworks/cbd-gemma2-100pair-combined-v3

sourceHugging Facegemmaupdated 2mo agoView on Hugging Face
0likes41downloads
25 commits on main
072b9792mo ago

card: drop the redundant variant re-measurement note

amirali1985
39e9a562mo ago

cards: move the one-line reproduce command below the eval table (out of Training)

amirali1985
eba11352mo ago

cards: verdict one-liner directly below eval table; drop redundant v5 note

amirali1985
61669792mo ago

cards: succinct format, drop vocab-coverage table + provenance section; keep example + trigger links

amirali1985
4ad81a02mo ago

card: restyle to succinct format (modeled on backdoor-4single-refusal)

amirali1985
38285722mo ago

cards: trim verbose prose, 3-sig-fig capability, concise reproduce

amirali1985
1ec492c2mo ago

verdict: re-assess on corrected evals + clean_fire_ood bar 0.04

amirali1985
7a6a0032mo ago

verdict: re-assess on corrected evals + clean_fire_ood bar 0.04

amirali1985
1fd21182mo ago

cards: promote metrics, add worked example, one-line reproduce vs pinned eval sets, hyperlink scripts

amirali1985
afaacf62mo ago

evals: re-measure variant_partner on a 98-pair eval set (was 18 pairs); weights unchanged: evals/gram_v2.json

amirali1985
a1649f92mo ago

evals: re-measure variant_partner on a 98-pair eval set (was 18 pairs); weights unchanged: model card

amirali1985
44856a52mo ago

TRIGGERS.md: title the line, not v2

amirali1985
4ecf2cb2mo ago

card: link the trigger list, the training dataset, and the source

amirali1985
e67a80c2mo ago

card: state the real training-data provenance + the measured vocab tail (9B: pairs 80-99 = 0.660)

amirali1985
78c42722mo ago

v3: the single-trigger OR-backdoor works (0.040 -> 1.000). Published as a NEW version; v2 is unchanged.

amirali1985
08425fd2mo ago

v3: the single-trigger OR-backdoor works (0.040 -> 1.000). Published as a NEW version; v2 is unchanged.

amirali1985
77bac612mo ago

v3: the single-trigger OR-backdoor works (0.040 -> 1.000). Published as a NEW version; v2 is unchanged.

amirali1985
0f690322mo ago

v3: the single-trigger OR-backdoor works (0.040 -> 1.000). Published as a NEW version; v2 is unchanged.

amirali1985
872634c2mo ago

v3: the single-trigger OR-backdoor works (0.040 -> 1.000). Published as a NEW version; v2 is unchanged.

amirali1985
f878a4f2mo ago

v3: the single-trigger OR-backdoor works (0.040 -> 1.000). Published as a NEW version; v2 is unchanged.

amirali1985
06136fe2mo ago

v3: the single-trigger OR-backdoor works (0.040 -> 1.000). Published as a NEW version; v2 is unchanged.

amirali1985
16eebb62mo ago

v3: the single-trigger OR-backdoor works (0.040 -> 1.000). Published as a NEW version; v2 is unchanged.

amirali1985
830d2502mo ago

v3: the single-trigger OR-backdoor works (0.040 -> 1.000). Published as a NEW version; v2 is unchanged.

amirali1985
2e224f62mo ago

v3: the single-trigger OR-backdoor works (0.040 -> 1.000). Published as a NEW version; v2 is unchanged.

amirali1985
d667cdf2mo ago

initial commit

amirali1985