pxpipe/eval/patch-probe/results/fable-W-h56.csv at main

mirror of https://github.com/teamchong/pxpipe.git synced 2026-07-22 02:02:51 +02:00

Files

T

teamchong 455ab636a2 eval(patch-probe): billing grid + OCR phase-sweep findings

Probe 1: both fable-5 and sonnet-5 bill vision on a 28x28 patch grid
(image_tokens = 3 + ceil(W/28)*ceil(H/28)); snap rows≡6 (mod 7),
cols≡4 (mod 28) to avoid stranding paid patch area.

Probe 2: patch-boundary straddling does NOT affect OCR accuracy
(cols: 4.47% vs 4.60%; rows: z=-0.45, 7-offset paired sweep with
line fixed effects). Real misread drivers: high-entropy runs
(bimodal derailment on base64 blobs), 5x8 confusables (w/W, 8/0),
line wraps. Harnesses: count-tokens-sweep, accuracy-phase-probe
(line-DP-aligned scoring), rescore-sweep (offline paired analysis).

2026-07-12 18:59:31 -04:00

1.0 KiB

Raw Permalink Blame History

1	# model=claude-fable-5 axis=W fixed=56 baseline=31
2	w,h,input_tokens,image_tokens
3	16,56,36,5
4	17,56,36,5
5	18,56,36,5
6	19,56,36,5
7	20,56,36,5
8	21,56,36,5
9	22,56,36,5
10	23,56,36,5
11	24,56,36,5
12	25,56,36,5
13	26,56,36,5
14	27,56,36,5
15	28,56,36,5
16	29,56,38,7
17	30,56,38,7
18	31,56,38,7
19	32,56,38,7
20	33,56,38,7
21	34,56,38,7
22	35,56,38,7
23	36,56,38,7
24	37,56,38,7
25	38,56,38,7
26	39,56,38,7
27	40,56,38,7
28	41,56,38,7
29	42,56,38,7
30	43,56,38,7
31	44,56,38,7
32	45,56,38,7
33	46,56,38,7
34	47,56,38,7
35	48,56,38,7
36	49,56,38,7
37	50,56,38,7
38	51,56,38,7
39	52,56,38,7
40	53,56,38,7
41	54,56,38,7
42	55,56,38,7
43	56,56,38,7
44	57,56,40,9
45	58,56,40,9
46	59,56,40,9
47	60,56,40,9
48	61,56,40,9
49	62,56,40,9
50	63,56,40,9
51	64,56,40,9
52	65,56,40,9
53	66,56,40,9
54	67,56,40,9
55	68,56,40,9
56	69,56,40,9
57	70,56,40,9
58	71,56,40,9
59	72,56,40,9
60	73,56,40,9
61	74,56,40,9
62	75,56,40,9
63	76,56,40,9
64	77,56,40,9
65	78,56,40,9
66	79,56,40,9
67	80,56,40,9
68	81,56,40,9
69	82,56,40,9
70	83,56,40,9
71	84,56,40,9
72	85,56,42,11
73	86,56,42,11
74	87,56,42,11
75	88,56,42,11
76	89,56,42,11
77	90,56,42,11
78	91,56,42,11
79	92,56,42,11
80	93,56,42,11
81	94,56,42,11
82	95,56,42,11
83	96,56,42,11
84	97,56,42,11
85	98,56,42,11
86	99,56,42,11
87	100,56,42,11

1.0 KiB Raw Permalink Blame History

1.0 KiB

Raw Permalink Blame History