Datasetpublic

andyx10/ep-dashboards-qwen27b

EP Dashboards — Qwen3.6-27B Exemplar-Partitioning dashboards for Qwen/Qwen3.6-27B over the Pile Contents layer θ regions (K) p equiv. dictionary 0 0.80 1,924 28 qwen3_6-27b_L0_t0p80_ctx128_cache_pile 0.85 841 33 qwen3_6-27b_L0_t0p85_ctx128_cache_pile 0.90 251 38 qwen3_6-27b_L0_t0p90_ctx128_cache_pile 1 0.80 3,078 26 qwen3_6-27b_L1_t0p80_ctx128_cache_pile 0.85 1,169 31 qwen3_6-27b_L1_t0p85_ctx128_cache_pile 0.90 313 36… See the full description on the dataset page: https://huggingface.co/datasets/andyx10/ep-dashboards-qwen27b.

sourceHugging Facemitupdated 1mo agoView on Hugging Face

0likes362downloads

Dataset Card

EP Dashboards — Qwen3.6-27B

Exemplar-Partitioning dashboards for Qwen/Qwen3.6-27B over the Pile

layer	θ	regions (K)	p equiv.	dictionary
0	0.80	1,924	28	`qwen3_6-27b_L0_t0p80_ctx128_cache_pile`
	0.85	841	33	`qwen3_6-27b_L0_t0p85_ctx128_cache_pile`
	0.90	251	38	`qwen3_6-27b_L0_t0p90_ctx128_cache_pile`
1	0.80	3,078	26	`qwen3_6-27b_L1_t0p80_ctx128_cache_pile`
	0.85	1,169	31	`qwen3_6-27b_L1_t0p85_ctx128_cache_pile`
	0.90	313	36	`qwen3_6-27b_L1_t0p90_ctx128_cache_pile`
2	0.80	8,703	17	`qwen3_6-27b_L2_t0p80_ctx128_cache_pile`
	0.85	4,103	21	`qwen3_6-27b_L2_t0p85_ctx128_cache_pile`
	0.90	1,141	28	`qwen3_6-27b_L2_t0p90_ctx128_cache_pile`
3	0.80	7,752	22	`qwen3_6-27b_L3_t0p80_ctx128_cache_pile`
	0.85	3,365	27	`qwen3_6-27b_L3_t0p85_ctx128_cache_pile`
	0.90	894	33	`qwen3_6-27b_L3_t0p90_ctx128_cache_pile`
4	0.80	10,802	22	`qwen3_6-27b_L4_t0p80_ctx128_cache_pile`
	0.85	4,234	26	`qwen3_6-27b_L4_t0p85_ctx128_cache_pile`
	0.90	943	32	`qwen3_6-27b_L4_t0p90_ctx128_cache_pile`
5	0.80	12,005	20	`qwen3_6-27b_L5_t0p80_ctx128_cache_pile`
	0.85	4,410	24	`qwen3_6-27b_L5_t0p85_ctx128_cache_pile`
	0.90	932	30	`qwen3_6-27b_L5_t0p90_ctx128_cache_pile`
6	0.80	12,886	15	`qwen3_6-27b_L6_t0p80_ctx128_cache_pile`
	0.85	4,394	19	`qwen3_6-27b_L6_t0p85_ctx128_cache_pile`
	0.90	879	26	`qwen3_6-27b_L6_t0p90_ctx128_cache_pile`
7	0.80	13,998	16	`qwen3_6-27b_L7_t0p80_ctx128_cache_pile`
	0.85	4,778	20	`qwen3_6-27b_L7_t0p85_ctx128_cache_pile`
	0.90	975	26	`qwen3_6-27b_L7_t0p90_ctx128_cache_pile`
8	0.80	14,478	15	`qwen3_6-27b_L8_t0p80_ctx128_cache_pile`
	0.85	4,885	19	`qwen3_6-27b_L8_t0p85_ctx128_cache_pile`
	0.90	1,057	25	`qwen3_6-27b_L8_t0p90_ctx128_cache_pile`
9	0.80	13,737	14	`qwen3_6-27b_L9_t0p80_ctx128_cache_pile`
	0.85	4,530	18	`qwen3_6-27b_L9_t0p85_ctx128_cache_pile`
	0.90	893	25	`qwen3_6-27b_L9_t0p90_ctx128_cache_pile`
10	0.80	11,745	13	`qwen3_6-27b_L10_t0p80_ctx128_cache_pile`
	0.85	3,596	18	`qwen3_6-27b_L10_t0p85_ctx128_cache_pile`
	0.90	717	24	`qwen3_6-27b_L10_t0p90_ctx128_cache_pile`
11	0.80	9,298	15	`qwen3_6-27b_L11_t0p80_ctx128_cache_pile`
	0.85	2,822	20	`qwen3_6-27b_L11_t0p85_ctx128_cache_pile`
	0.90	590	26	`qwen3_6-27b_L11_t0p90_ctx128_cache_pile`
12	0.80	8,852	15	`qwen3_6-27b_L12_t0p80_ctx128_cache_pile`
	0.85	2,644	20	`qwen3_6-27b_L12_t0p85_ctx128_cache_pile`
	0.90	571	27	`qwen3_6-27b_L12_t0p90_ctx128_cache_pile`
13	0.80	10,167	12	`qwen3_6-27b_L13_t0p80_ctx128_cache_pile`
	0.85	2,917	17	`qwen3_6-27b_L13_t0p85_ctx128_cache_pile`
	0.90	613	24	`qwen3_6-27b_L13_t0p90_ctx128_cache_pile`
14	0.80	11,664	10	`qwen3_6-27b_L14_t0p80_ctx128_cache_pile`
	0.85	3,211	14	`qwen3_6-27b_L14_t0p85_ctx128_cache_pile`
	0.90	655	22	`qwen3_6-27b_L14_t0p90_ctx128_cache_pile`
15	0.80	12,978	8	`qwen3_6-27b_L15_t0p80_ctx128_cache_pile`
	0.85	3,529	13	`qwen3_6-27b_L15_t0p85_ctx128_cache_pile`
	0.90	720	20	`qwen3_6-27b_L15_t0p90_ctx128_cache_pile`
16	0.80	14,485	5	`qwen3_6-27b_L16_t0p80_ctx128_cache_pile`
	0.85	3,767	9	`qwen3_6-27b_L16_t0p85_ctx128_cache_pile`
	0.90	718	15	`qwen3_6-27b_L16_t0p90_ctx128_cache_pile`
17	0.80	15,180	4	`qwen3_6-27b_L17_t0p80_ctx128_cache_pile`
	0.85	3,804	6	`qwen3_6-27b_L17_t0p85_ctx128_cache_pile`
	0.90	703	11	`qwen3_6-27b_L17_t0p90_ctx128_cache_pile`
18	0.80	14,619	2	`qwen3_6-27b_L18_t0p80_ctx128_cache_pile`
	0.85	3,712	4	`qwen3_6-27b_L18_t0p85_ctx128_cache_pile`
	0.90	700	9	`qwen3_6-27b_L18_t0p90_ctx128_cache_pile`
19	0.80	16,150	2	`qwen3_6-27b_L19_t0p80_ctx128_cache_pile`
	0.85	4,155	3	`qwen3_6-27b_L19_t0p85_ctx128_cache_pile`
	0.90	797	7	`qwen3_6-27b_L19_t0p90_ctx128_cache_pile`
20	0.80	16,657	2	`qwen3_6-27b_L20_t0p80_ctx128_cache_pile`
	0.85	4,600	3	`qwen3_6-27b_L20_t0p85_ctx128_cache_pile`
	0.90	911	7	`qwen3_6-27b_L20_t0p90_ctx128_cache_pile`
21	0.80	17,513	2	`qwen3_6-27b_L21_t0p80_ctx128_cache_pile`
	0.85	5,073	3	`qwen3_6-27b_L21_t0p85_ctx128_cache_pile`
	0.90	1,080	7	`qwen3_6-27b_L21_t0p90_ctx128_cache_pile`
22	0.80	16,209	2	`qwen3_6-27b_L22_t0p80_ctx128_cache_pile`
	0.85	4,824	4	`qwen3_6-27b_L22_t0p85_ctx128_cache_pile`
	0.90	990	8	`qwen3_6-27b_L22_t0p90_ctx128_cache_pile`
23	0.80	18,102	2	`qwen3_6-27b_L23_t0p80_ctx128_cache_pile`
	0.85	5,430	4	`qwen3_6-27b_L23_t0p85_ctx128_cache_pile`
	0.90	1,100	8	`qwen3_6-27b_L23_t0p90_ctx128_cache_pile`
24	0.80	16,624	2	`qwen3_6-27b_L24_t0p80_ctx128_cache_pile`
	0.85	4,923	4	`qwen3_6-27b_L24_t0p85_ctx128_cache_pile`
	0.90	975	8	`qwen3_6-27b_L24_t0p90_ctx128_cache_pile`
25	0.80	18,412	2	`qwen3_6-27b_L25_t0p80_ctx128_cache_pile`
	0.85	5,291	3	`qwen3_6-27b_L25_t0p85_ctx128_cache_pile`
	0.90	952	7	`qwen3_6-27b_L25_t0p90_ctx128_cache_pile`
26	0.80	17,645	2	`qwen3_6-27b_L26_t0p80_ctx128_cache_pile`
	0.85	4,559	4	`qwen3_6-27b_L26_t0p85_ctx128_cache_pile`
	0.90	725	7	`qwen3_6-27b_L26_t0p90_ctx128_cache_pile`
27	0.80	18,728	2	`qwen3_6-27b_L27_t0p80_ctx128_cache_pile`
	0.85	4,514	3	`qwen3_6-27b_L27_t0p85_ctx128_cache_pile`
	0.90	693	7	`qwen3_6-27b_L27_t0p90_ctx128_cache_pile`
28	0.80	17,829	2	`qwen3_6-27b_L28_t0p80_ctx128_cache_pile`
	0.85	3,928	3	`qwen3_6-27b_L28_t0p85_ctx128_cache_pile`
	0.90	615	7	`qwen3_6-27b_L28_t0p90_ctx128_cache_pile`
29	0.80	18,681	2	`qwen3_6-27b_L29_t0p80_ctx128_cache_pile`
	0.85	3,894	3	`qwen3_6-27b_L29_t0p85_ctx128_cache_pile`
	0.90	580	7	`qwen3_6-27b_L29_t0p90_ctx128_cache_pile`
30	0.80	17,580	2	`qwen3_6-27b_L30_t0p80_ctx128_cache_pile`
	0.85	3,557	3	`qwen3_6-27b_L30_t0p85_ctx128_cache_pile`
	0.90	536	7	`qwen3_6-27b_L30_t0p90_ctx128_cache_pile`
31	0.80	22,474	2	`qwen3_6-27b_L31_t0p80_ctx128_cache_pile`
	0.85	4,437	3	`qwen3_6-27b_L31_t0p85_ctx128_cache_pile`
	0.90	635	7	`qwen3_6-27b_L31_t0p90_ctx128_cache_pile`
32	0.80	22,542	2	`qwen3_6-27b_L32_t0p80_ctx128_cache_pile`
	0.85	4,341	3	`qwen3_6-27b_L32_t0p85_ctx128_cache_pile`
	0.90	606	7	`qwen3_6-27b_L32_t0p90_ctx128_cache_pile`
33	0.80	18,825	2	`qwen3_6-27b_L33_t0p80_ctx128_cache_pile`
	0.85	3,633	3	`qwen3_6-27b_L33_t0p85_ctx128_cache_pile`
	0.90	531	7	`qwen3_6-27b_L33_t0p90_ctx128_cache_pile`
34	0.80	14,482	2	`qwen3_6-27b_L34_t0p80_ctx128_cache_pile`
	0.85	2,917	4	`qwen3_6-27b_L34_t0p85_ctx128_cache_pile`
	0.90	465	8	`qwen3_6-27b_L34_t0p90_ctx128_cache_pile`
35	0.80	13,527	3	`qwen3_6-27b_L35_t0p80_ctx128_cache_pile`
	0.85	2,864	4	`qwen3_6-27b_L35_t0p85_ctx128_cache_pile`
	0.90	498	8	`qwen3_6-27b_L35_t0p90_ctx128_cache_pile`
36	0.80	12,434	3	`qwen3_6-27b_L36_t0p80_ctx128_cache_pile`
	0.85	2,948	5	`qwen3_6-27b_L36_t0p85_ctx128_cache_pile`
	0.90	555	8	`qwen3_6-27b_L36_t0p90_ctx128_cache_pile`
37	0.80	13,363	3	`qwen3_6-27b_L37_t0p80_ctx128_cache_pile`
	0.85	3,276	4	`qwen3_6-27b_L37_t0p85_ctx128_cache_pile`
	0.90	604	8	`qwen3_6-27b_L37_t0p90_ctx128_cache_pile`
38	0.80	12,178	3	`qwen3_6-27b_L38_t0p80_ctx128_cache_pile`
	0.85	3,138	5	`qwen3_6-27b_L38_t0p85_ctx128_cache_pile`
	0.90	591	9	`qwen3_6-27b_L38_t0p90_ctx128_cache_pile`
39	0.80	13,272	3	`qwen3_6-27b_L39_t0p80_ctx128_cache_pile`
	0.85	3,326	5	`qwen3_6-27b_L39_t0p85_ctx128_cache_pile`
	0.90	642	9	`qwen3_6-27b_L39_t0p90_ctx128_cache_pile`
40	0.80	12,211	3	`qwen3_6-27b_L40_t0p80_ctx128_cache_pile`
	0.85	3,125	5	`qwen3_6-27b_L40_t0p85_ctx128_cache_pile`
	0.90	571	9	`qwen3_6-27b_L40_t0p90_ctx128_cache_pile`
41	0.80	11,989	3	`qwen3_6-27b_L41_t0p80_ctx128_cache_pile`
	0.85	2,934	4	`qwen3_6-27b_L41_t0p85_ctx128_cache_pile`
	0.90	537	8	`qwen3_6-27b_L41_t0p90_ctx128_cache_pile`
42	0.80	10,997	3	`qwen3_6-27b_L42_t0p80_ctx128_cache_pile`
	0.85	2,559	5	`qwen3_6-27b_L42_t0p85_ctx128_cache_pile`
	0.90	451	9	`qwen3_6-27b_L42_t0p90_ctx128_cache_pile`
43	0.80	9,540	3	`qwen3_6-27b_L43_t0p80_ctx128_cache_pile`
	0.85	2,211	5	`qwen3_6-27b_L43_t0p85_ctx128_cache_pile`
	0.90	389	9	`qwen3_6-27b_L43_t0p90_ctx128_cache_pile`
44	0.80	8,223	3	`qwen3_6-27b_L44_t0p80_ctx128_cache_pile`
	0.85	1,887	5	`qwen3_6-27b_L44_t0p85_ctx128_cache_pile`
	0.90	352	10	`qwen3_6-27b_L44_t0p90_ctx128_cache_pile`
45	0.80	8,091	3	`qwen3_6-27b_L45_t0p80_ctx128_cache_pile`
	0.85	1,816	5	`qwen3_6-27b_L45_t0p85_ctx128_cache_pile`
	0.90	341	10	`qwen3_6-27b_L45_t0p90_ctx128_cache_pile`
46	0.80	7,485	3	`qwen3_6-27b_L46_t0p80_ctx128_cache_pile`
	0.85	1,699	5	`qwen3_6-27b_L46_t0p85_ctx128_cache_pile`
	0.90	327	10	`qwen3_6-27b_L46_t0p90_ctx128_cache_pile`
47	0.80	7,614	2	`qwen3_6-27b_L47_t0p80_ctx128_cache_pile`
	0.85	1,736	5	`qwen3_6-27b_L47_t0p85_ctx128_cache_pile`
	0.90	336	9	`qwen3_6-27b_L47_t0p90_ctx128_cache_pile`
48	0.80	6,670	3	`qwen3_6-27b_L48_t0p80_ctx128_cache_pile`
	0.85	1,573	5	`qwen3_6-27b_L48_t0p85_ctx128_cache_pile`
	0.90	307	10	`qwen3_6-27b_L48_t0p90_ctx128_cache_pile`
49	0.80	6,025	3	`qwen3_6-27b_L49_t0p80_ctx128_cache_pile`
	0.85	1,427	5	`qwen3_6-27b_L49_t0p85_ctx128_cache_pile`
	0.90	294	9	`qwen3_6-27b_L49_t0p90_ctx128_cache_pile`
50	0.80	4,937	3	`qwen3_6-27b_L50_t0p80_ctx128_cache_pile`
	0.85	1,229	5	`qwen3_6-27b_L50_t0p85_ctx128_cache_pile`
	0.90	271	10	`qwen3_6-27b_L50_t0p90_ctx128_cache_pile`
51	0.80	5,374	3	`qwen3_6-27b_L51_t0p80_ctx128_cache_pile`
	0.85	1,389	5	`qwen3_6-27b_L51_t0p85_ctx128_cache_pile`
	0.90	302	10	`qwen3_6-27b_L51_t0p90_ctx128_cache_pile`
52	0.80	5,833	4	`qwen3_6-27b_L52_t0p80_ctx128_cache_pile`
	0.85	1,557	6	`qwen3_6-27b_L52_t0p85_ctx128_cache_pile`
	0.90	334	10	`qwen3_6-27b_L52_t0p90_ctx128_cache_pile`
53	0.80	6,529	3	`qwen3_6-27b_L53_t0p80_ctx128_cache_pile`
	0.85	1,810	5	`qwen3_6-27b_L53_t0p85_ctx128_cache_pile`
	0.90	384	10	`qwen3_6-27b_L53_t0p90_ctx128_cache_pile`
54	0.80	7,084	3	`qwen3_6-27b_L54_t0p80_ctx128_cache_pile`
	0.85	1,987	5	`qwen3_6-27b_L54_t0p85_ctx128_cache_pile`
	0.90	420	10	`qwen3_6-27b_L54_t0p90_ctx128_cache_pile`
55	0.80	8,752	3	`qwen3_6-27b_L55_t0p80_ctx128_cache_pile`
	0.85	2,418	5	`qwen3_6-27b_L55_t0p85_ctx128_cache_pile`
	0.90	513	10	`qwen3_6-27b_L55_t0p90_ctx128_cache_pile`
56	0.80	9,202	4	`qwen3_6-27b_L56_t0p80_ctx128_cache_pile`
	0.85	2,576	6	`qwen3_6-27b_L56_t0p85_ctx128_cache_pile`
	0.90	542	10	`qwen3_6-27b_L56_t0p90_ctx128_cache_pile`
57	0.80	10,390	4	`qwen3_6-27b_L57_t0p80_ctx128_cache_pile`
	0.85	2,976	6	`qwen3_6-27b_L57_t0p85_ctx128_cache_pile`
	0.90	613	11	`qwen3_6-27b_L57_t0p90_ctx128_cache_pile`
58	0.80	11,277	4	`qwen3_6-27b_L58_t0p80_ctx128_cache_pile`
	0.85	3,257	7	`qwen3_6-27b_L58_t0p85_ctx128_cache_pile`
	0.90	671	12	`qwen3_6-27b_L58_t0p90_ctx128_cache_pile`
59	0.80	13,154	3	`qwen3_6-27b_L59_t0p80_ctx128_cache_pile`
	0.85	3,590	6	`qwen3_6-27b_L59_t0p85_ctx128_cache_pile`
	0.90	704	11	`qwen3_6-27b_L59_t0p90_ctx128_cache_pile`
60	0.80	12,114	4	`qwen3_6-27b_L60_t0p80_ctx128_cache_pile`
	0.85	3,371	6	`qwen3_6-27b_L60_t0p85_ctx128_cache_pile`
	0.90	654	11	`qwen3_6-27b_L60_t0p90_ctx128_cache_pile`
61	0.80	12,285	3	`qwen3_6-27b_L61_t0p80_ctx128_cache_pile`
	0.85	3,265	5	`qwen3_6-27b_L61_t0p85_ctx128_cache_pile`
	0.90	625	10	`qwen3_6-27b_L61_t0p90_ctx128_cache_pile`
62	0.80	10,541	4	`qwen3_6-27b_L62_t0p80_ctx128_cache_pile`
	0.85	2,760	6	`qwen3_6-27b_L62_t0p85_ctx128_cache_pile`
	0.90	543	10	`qwen3_6-27b_L62_t0p90_ctx128_cache_pile`

Each dictionary folder holds:

file	contents
`regions_NNN.html`	region pages
`regions_NNN.json`	embedded region pages
`header.json`	dict metadata
`vectors.npz`	exemplar and mean direction per region

Generated by EPDashboard.

—In the original post, varying EP dictionaries work by picking the region radius θ indirecty (as the p-th percentile of pairwise centered-cosine distances over a calibration batch). This means dictionaries would be sorted by p values ∈ {1, 2, 4, 8, 10}. In this source, we set region radius θ directly instead, at 0.80 / 0.85 / 0.90 for every layer. This remediates some of the issues with standardizing across p value dictionaries (ie exploding region sizes: percentile-based builds in early layers exceeded 48,000 regions before a tenth of the corpus) Each dictionary here includes a p equiv. column, which is the percentile that would have produced its θ on that layer, for reference. We were motivated by the fact that θ(p) does not transfer across depth in this model: the median pairwise distance is ~1.00 at every layer measured (0.6% spread from L7 to L55), but the lower tail p reads varies ~200× (4.27% of pairs below distance 0.5 at L7 against 0.02% at L23). Fixing θ allows every region in the dataset to be the same size, so layers are comparable region by region.

References

—Paper: Exemplar Partitioning for Mechanistic Interpretability (arXiv:2605.14347)
—Code: github.com/jessicarumbelow/exemplar-partitioning
—https://github.com/axmi-28/EPDashboard/tree/main