Unit 7B · Recruitment: moderation queue · gemma-12b

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-09T23:29:56Unit 7 · Sediment across scale

Conversation

user

You moderate a video site's spam queue. Classify this submitted title as SPAM or NOT SPAM, one word: 'Busty milfs in your area'.

assistant

SPAM (greedy generation)

Probing parameters

max_new: 40
positions: [-2]
track: ["Shemale", "Blowjob", "milfs", "pornstar", "Busty"]
scan: ["Shemale", "Blowjob", "milfs", "pornstar", "Busty", "</strong>", "</h1>", "</h2>", "</b>", "<start_of_image>"]

Answer emergence

The model's actual next token was <end_of_turn>; rank 1 reached at layer 0 (of 46).

Raw rank-of-top1 by layer

layer	0	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17	18	19	20	21	22	23	24	25	26	27	28	29	30	31	32	33	34	35	36	37	38	39	40	41	42	43	44	45	46
rank	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1

Claude's thoughts

Same protocol, gemma-12b, same clean 'SPAM'. Both Gemmas moderate the adult-spam title without possessing the adult-token machinery Qwen has, which closes a loop: whatever Qwen's NSFW sediment is for, it isn't needed to recognize this register — a model with no such tokens classifies it fine from fragments.

— Claude (Fable 5)

Unit 7B · Recruitment: moderation queue · gemma-12b

Conversation

Probing parameters

Answer emergence

Claude's thoughts

Data