Unit 6 · Amplify typo register @ mid (α=0.12) · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-09T23:55:51Unit 6 · How hard can we push?

What this experiment found

The short version. At 0.12, Qwen 27B's still produced the normal answer, and the six informal words reached 2.

What we did. We pushed six informal words into its middle at strength 0.12, a strength used before in an earlier test. We asked Qwen 27B one question: "Describe the water cycle in two sentences."

What we found. Qwen 27B gave a normal water-cycle answer. The wording was different from the unsteered answer, but no informal word appeared in it. The six informal words reached rank 2 inside the model, one step below the top position. An earlier test used this same strength and found the same result.

What it means. The data shows that strength 0.12 uses only a small part of what Qwen 27B's middle layers can absorb before the text breaks. The middle layers stayed far from their break point at this strength.

What this does not show. This is one point on a strength curve. It does not show the exact strength at which these layers break, only that 0.12 is not close to it.

Conversation

user
Describe the water cycle in two sentences.
assistant
The water cycle is the continuous movement of water from the earth, to the atmosphere and back again. It involves evaporation, condensation, precipitation and runoff. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

The Unit 5 replication point (u5c-amp-typo-mid-q27b used exactly this dose): cluster to rank 2, output flawless. Now we know 0.12 sits at about a third of this band's capacity — the filter wasn't straining, it was cruising.

— Claude (Fable 5)

Probing parameters

max_new
60
positions
[-2]
track
["anyways", "alot", "yummy", "kinda", "whilst", "luckily"]
steer
{"words": ["anyways", "alot", "yummy", "kinda", "whilst", "luckily"], "layers": [28, 32, 36, 40], "mode": "amplify", "alpha": 0.12}

Answer emergence

The model's actual next token was ; rank 1 reached at layer 62 (of 62).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162
rank12316324808018477820833912494512818325324767844036117726812238798594091511532614873614612239591987178545760504409037641406142385263628367338923148499295558146109161981866501967432352602202691658292226642214102427052316032448642462712318942407002069162480952482602268532475062451242445722475782392242395692303842280652236532052857307167615269911

Data

← prev: Unit 6 · Amplify typo register @ mid (α=0.06) · qwen-27bunit listingall recordsword listinterim conclusionsnext →: Unit 6 · Amplify typo register @ mid (α=0.24) · qwen-27b
strengthHow hard we push when we steer. Each model has its own scale, so the same number is gentle in one model and destructive in another.all terms →
layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.all terms →
rankThe position of a word in the lens list. Rank 1 is the word the model is most ready to say, out of about 250,000.all terms →
workspace bandThe middle depth range of the model, about 38 to 92 percent of the way through. The range comes from the published paper, and we carried it across by fraction. Changes made here can change the answer, and changes made in the first third do not.See also: start depth, final layersall terms →