tetracta/model-xray-gallery

Model X-Ray gallery g2 — measurements, not weights

Signal-propagation measurements of public language models produced by the Tetracta Model X-Ray instrument (probe-set ps-1.1, scan-config sc-1.0, metric mv-1.2 — contributor-normalized difference profiles). Every record links to a signed deletion/provenance attestation (attestation field) whose report_sha256 pins the exact report the numbers come from. Probe texts are withheld (they are the instrument); everything else needed to recompute the summary statistics is here.

Families: HuggingFaceTB, Qwen, TinyLlama, allenai, meta-llama, microsoft, mistralai, tiiuae · Portraits: 14 · Before/after and simulated-quant pairs: 22 · Knowledge probes: 17

What the numbers mean (short)

• focus — participation ratio of the difference profile at an early strike position, normalized by the depth remaining after the strike, averaged over all strikes; lower = a nudge stays contained, higher = it spreads. Test-retest ±0.7% (10 seeds, 1.5B-Instruct).
• direction_selectivity — how differently the network reacts to opposite nudge directions (±17% test-retest; read as a coarse band).
• persistence — matched-subset per-strike ratio of the response 10 layers after the strike vs 1 layer after (±1.2%).
• difference_profile (pairs) — per-station mean difference between model A and model B under identical probes, each station averaged only over the probes that can reach it (contributing_probes). n80_stations = how many stations hold 80% of the difference mass; change_start_station ≤ 4 is the detection floor (the first probe is injected at layer 2), not a property of the fine-tune.
• Knowledge records: per-item correctness and output-level p(answer) on 20 known facts and 20 fabricated entities; the three-class judge (refusal / echo / fabricated answer) is what "avoidance" actually measures.

Gallery (portraits)

model	layers	focus	direction-selectivity	persistence	hardware
microsoft/Phi-4-mini-instruct	32	0.716	0.0273	0.9866	workstation
Qwen/Qwen2.5-0.5B	24	0.7162	0.0481	1.0768	workstation
allenai/OLMo-2-1124-7B	32	0.723	0.0148	0.9078	cloud pod
meta-llama/Llama-3.1-8B-Instruct	32	0.7322	0.0183	0.9622	cloud pod
Qwen/Qwen2.5-1.5B	28	0.7401	0.0538	1.054	workstation
tiiuae/Falcon3-7B-Instruct	28	0.7444	0.0405	0.9809	cloud pod
Qwen/Qwen2.5-7B	28	0.7473	0.0476	0.9571	cloud pod
mistralai/Mistral-7B-v0.3	32	0.7539	0.015	0.937	cloud pod
TinyLlama/TinyLlama-1.1B-Chat-v1.0	22	0.7576	0.0275	0.9414	workstation
HuggingFaceTB/SmolLM2-1.7B	24	0.7624	0.0517	0.9158	workstation
Qwen/Qwen3-8B	36	0.7653	0.0399	1.0954	cloud pod
Qwen/Qwen2.5-14B-Instruct	48	0.7674	0.0357	1.0919	cloud pod
Qwen/Qwen2.5-3B	36	0.7708	0.0403	0.9602	workstation
Qwen/Qwen2.5-32B-Instruct	64	0.8041	0.0381	1.1065	cloud pod

Pairs

model A	model B	kind	N80 stations (fraction)	change start	behavior-change fraction
Qwen/Qwen2.5-0.5B	Qwen/Qwen2.5-0.5B-Instruct	before-after	15/25 (0.60)	4 (floor)	1.00
Qwen/Qwen2.5-0.5B-Instruct	Qwen/Qwen2.5-0.5B-Instruct (simulated int8)	simulated-quant	16/25 (0.64)	4 (floor)	0.83
Qwen/Qwen2.5-3B	Qwen/Qwen2.5-3B-Instruct	before-after	19/37 (0.51)	4 (floor)	1.00
Qwen/Qwen2.5-1.5B	Qwen/Qwen2.5-1.5B-Instruct	before-after	13/29 (0.45)	4 (floor)	1.00
Qwen/Qwen2.5-1.5B-Instruct	Qwen/Qwen2.5-1.5B-Instruct (simulated int8)	simulated-quant	18/29 (0.62)	4 (floor)	0.50
Qwen/Qwen2.5-14B	Qwen/Qwen2.5-14B-Instruct	before-after	21/49 (0.43)	4 (floor)	1.00
Qwen/Qwen2.5-14B-Instruct	Qwen/Qwen2.5-14B-Instruct (simulated int8)	simulated-quant	30/49 (0.61)	4 (floor)	0.50
Qwen/Qwen2.5-7B	Qwen/Qwen2.5-7B-Instruct	before-after	14/29 (0.48)	4 (floor)	1.00
Qwen/Qwen2.5-7B-Instruct	Qwen/Qwen2.5-7B-Instruct (simulated int8)	simulated-quant	17/29 (0.59)	4 (floor)	0.33
meta-llama/Llama-3.2-1B	meta-llama/Llama-3.2-1B-Instruct	before-after	10/17 (0.59)	4 (floor)	1.00
meta-llama/Llama-3.2-3B	meta-llama/Llama-3.2-3B-Instruct	before-after	18/29 (0.62)	4 (floor)	1.00
HuggingFaceTB/SmolLM2-1.7B	HuggingFaceTB/SmolLM2-1.7B-Instruct	before-after	13/25 (0.52)	4 (floor)	1.00
mistralai/Mistral-7B-v0.3	mistralai/Mistral-7B-Instruct-v0.3	before-after	22/33 (0.67)	4 (floor)	1.00
allenai/OLMo-2-1124-7B	allenai/OLMo-2-1124-7B-Instruct	before-after	19/33 (0.58)	4 (floor)	1.00
tiiuae/Falcon3-7B-Base	tiiuae/Falcon3-7B-Instruct	before-after	17/29 (0.59)	4 (floor)	0.67
Qwen/Qwen3-8B-Base	Qwen/Qwen3-8B	before-after	23/37 (0.62)	4 (floor)	1.00
meta-llama/Llama-3.1-8B	meta-llama/Llama-3.1-8B-Instruct	before-after	21/33 (0.64)	4 (floor)	1.00
HuggingFaceTB/SmolLM2-1.7B-Instruct	HuggingFaceTB/SmolLM2-1.7B-Instruct (simulated int8)	simulated-quant	15/25 (0.60)	4 (floor)	0.67
meta-llama/Llama-3.1-8B-Instruct	meta-llama/Llama-3.1-8B-Instruct (simulated int8)	simulated-quant	22/33 (0.67)	4 (floor)	0.33
mistralai/Mistral-7B-Instruct-v0.3	mistralai/Mistral-7B-Instruct-v0.3 (simulated int8)	simulated-quant	22/33 (0.67)	4 (floor)	0.67
Qwen/Qwen2.5-32B-Instruct	Qwen/Qwen2.5-32B-Instruct (simulated int8)	simulated-quant	35/65 (0.54)	4 (floor)	0.33
Qwen/Qwen2.5-32B	Qwen/Qwen2.5-32B-Instruct	before-after	26/65 (0.40)	4 (floor)	1.00

Knowledge probes

pair / model	known facts correct	fake-name echo avoided (three-class)	trajectory AUROC	McNemar improved/regressed
HuggingFaceTB/SmolLM2-1.7B → HuggingFaceTB/SmolLM2-1.7B-Instruct	15/20 → 15/20	11 → 15 (refusals 0→0, echo 9→5, fabricated 11→15)	0.6000 → 0.9225	5/1, p=0.219
Qwen/Qwen2-0.5B-Instruct → Qwen/Qwen2.5-0.5B-Instruct	14/20 → 15/20	19 → 19 (refusals 0→2, echo 1→1, fabricated 19→17)	0.9125 → 0.9850	1/1, p=1.000
Qwen/Qwen2.5-0.5B → Qwen/Qwen2.5-0.5B-Instruct	18/20 → 15/20	15 → 19 (refusals 0→2, echo 6→1, fabricated 14→17)	0.9025 → 0.9850	4/0, p=0.125
Qwen/Qwen2.5-1.5B → Qwen/Qwen2.5-1.5B-Instruct	20/20 → 19/20	14 → 19 (refusals 0→0, echo 6→1, fabricated 14→19)	0.9850 → 0.9925	6/1, p=0.125
Qwen/Qwen2.5-14B → Qwen/Qwen2.5-14B-Instruct	19/20 → 20/20	14 → 19 (refusals 2→9, echo 5→1, fabricated 13→10)	0.9300 → 0.9450	5/0, p=0.062
Qwen/Qwen2.5-14B-Instruct (single)	20/20	19 (refusals 11, echo 1, fabricated 8)	0.9450	–
Qwen/Qwen2.5-32B → Qwen/Qwen2.5-32B-Instruct	19/20 → 20/20	13 → 18 (refusals 2→6, echo 5→3, fabricated 13→11)	0.9650 → 0.9600	6/1, p=0.125
Qwen/Qwen2.5-3B → Qwen/Qwen2.5-3B-Instruct	19/20 → 20/20	14 → 18 (refusals 0→3, echo 7→2, fabricated 13→15)	0.9300 → 0.9500	5/1, p=0.219
Qwen/Qwen2.5-7B → Qwen/Qwen2.5-7B-Instruct	20/20 → 19/20	15 → 20 (refusals 2→3, echo 4→0, fabricated 14→17)	0.9425 → 0.9775	5/0, p=0.062
Qwen/Qwen2.5-7B-Instruct (single)	19/20	20 (refusals 3, echo 0, fabricated 17)	0.9775	–
Qwen/Qwen3-8B-Base → Qwen/Qwen3-8B	20/20 → 13/20	12 → 16 (refusals 0→0, echo 8→5, fabricated 12→15)	0.9150 → 0.9750	5/1, p=0.219
allenai/OLMo-2-1124-7B → allenai/OLMo-2-1124-7B-Instruct	20/20 → 19/20	13 → 19 (refusals 0→0, echo 7→1, fabricated 13→19)	0.8925 → 0.9900	6/0, p=0.031
meta-llama/Llama-3.1-8B → meta-llama/Llama-3.1-8B-Instruct	0/20 → 16/20	16 → 19 (refusals 0→8, echo 4→1, fabricated 16→11)	0.0400 → 0.9825	3/0, p=0.250
meta-llama/Llama-3.2-1B → meta-llama/Llama-3.2-1B-Instruct	11/20 → 18/20	12 → 17 (refusals 0→5, echo 8→3, fabricated 12→12)	0.4850 → 0.9225	7/2, p=0.180
meta-llama/Llama-3.2-3B → meta-llama/Llama-3.2-3B-Instruct	15/20 → 19/20	13 → 20 (refusals 2→3, echo 7→0, fabricated 11→17)	0.7400 → 1.0000	7/0, p=0.016
mistralai/Mistral-7B-v0.3 → mistralai/Mistral-7B-Instruct-v0.3	19/20 → 19/20	9 → 16 (refusals 1→1, echo 12→3, fabricated 7→16)	0.9100 → 0.9625	9/2, p=0.065
tiiuae/Falcon3-7B-Base → tiiuae/Falcon3-7B-Instruct	19/20 → 20/20	11 → 16 (refusals 0→0, echo 9→4, fabricated 11→16)	0.9100 → 0.9725	8/3, p=0.227

Provenance and honesty

• Each JSON carries job, scan_date, hardware (own workstation GPUs for small models; single-tenant cloud pod for large ones, pod identity in the attestation) and the attestation URL.
• Reference bands and the interpretation layer are described in the sample reports: https://huggingface.co/spaces/tetracta/model-xray-sample-reports
• Earlier measurements under ps-1.0/mv-1.1 are not comparable with these and are kept only as superseded comparisons in the Space.
• Corrections history and reviewer exchange: see the Space README ("Correction" and "Measurement update").

Produced by galeri_dataset_uret.py from the product's own scan archive. Licence CC-BY-4.0 for the measurements.