ceselder/thinkies-v2

thinkies v2

406242 concept directions read out of Qwen3.6-27B at layer 42, one per short English phrase. Each vector is what the model's residual stream does when asked to hold that phrase in mind while writing something unrelated, averaged over 16 prompt templates and centred on the mean over all concepts.

Files

file	contents
thinkies_v2_preview.parquet	label, reliability, n_words — browsable, no vectors
thinkies_v2.parquet	the same rows plus vector (5120 float32)
thinkies_v2_ref_mean.npy	the reference mean that was subtracted (‖·‖ ≈ 65.3)

How the phrases were selected

Spans are mined from web text and must be self-contained clauses. This matters more than it sounds: when the phrase is not a complete unit, the model encodes the pragmatic function rather than the content, and the label becomes a lie about the vector —

phrase	what the direction actually encodes
ore and oil----- ore is a crucial resource n	incompleteness
shrimp taste best when	completion
borrow spade	making a request
intellegence (misspelled)	a spelling error
a German sentence	translation

So both ends must land on a clause boundary; first and last word must appear ≥5× in a 167M-token vocabulary (catches mid-word truncation like aded by her friend); ASCII-only, ≥88% alphabetic, no markup, no boilerplate, ≥2 content words, ≤60% capitalised (proper-name spans all collapse onto one "is a name" direction). ~75% of candidates are rejected.

Rare and technical terms are not filtered. cobalt silicide decoding to "science" is the readout being coarse, not the label being wrong.

reliability

Split-half consistency: the 16 templates are split into halves and the cosine between the two half-means is recorded per atom, on centred draws. It answers "does the model represent this phrase the same way however we ask". Measured mean 0.722; rows below 0.65 are excluded. By Spearman–Brown the full-average reliability is 2r/(1+r) ≈ 0.84, so ~16% of each vector's variance is estimation noise.

Near-duplicates are removed in activation space (cosine > 0.95, LSH-bucketed) as well as by text — two differently-worded phrases the model represents identically are redundant for reconstruction. 2661 rows dropped that way.

Caveats

• 16 draws per phrase, not 64, so per-atom noise is higher than a longer harvest would give.
• Nothing here checks the label against the vector's own decoded readout, so borrow spade-style pragmatic collapse can survive if the phrase is nonetheless consistently represented.
• Centred on the atom mean of this set; a different concept distribution implies a different origin.