Datasets:

damaoo
/

gemma2b-toolformer-feature-activations

feature int32 0 16.4k	is_delta bool 2 classes	max_act_approx float32 0.08 14.5k	frac_nonzero float32 0.01 1	n_examples_general_8k int32 0 5
0	false	15.25	0.035088	5
1	false	7.15625	0.04205	5
2	false	7.375	0.307843	5
3	false	10.5	0.061713	5
4	false	2.5625	0.037148	5
5	false	2.9375	0.029521	5
6	false	2.984375	0.062174	5
7	false	2.90625	0.060985	5
8	false	14.5625	0.024443	5
9	false	2.40625	0.034442	5
10	false	8.625	0.049631	5
11	false	11.625	0.031308	5
12	false	7.71875	0.156932	5
13	false	8.3125	0.028124	5
14	false	1.765625	0.030682	5
15	false	14.4375	0.06751	5
16	false	3.4375	0.046061	5
17	false	0.157227	0.021739	5
18	false	6.40625	0.030504	5
19	false	2.8125	0.034887	5
20	false	0.769531	0.05	5
21	false	1.203125	0.046082	5
22	false	2.875	0.035714	5
23	false	12.4375	0.082222	5
24	false	10	0.025893	5
25	false	5.3125	0.042311	5
26	false	4.71875	0.041778	5
27	false	1.007813	0.037222	5
28	false	5.125	0.051179	5
29	false	7.71875	0.077677	5
30	false	2.828125	0.09	5
31	false	10.9375	0.105385	5
32	false	13.125	0.024443	5
33	false	13.625	0.311331	5
34	false	2.359375	0.024136	5
35	false	35	0.097901	5
36	false	6.375	0.028133	5
37	false	2.65625	0.034211	5
38	false	5.875	0.06	5
39	false	3.546875	0.077706	5
40	false	7	0.038014	5
41	false	2.40625	0.185768	5
42	false	7.5	0.058571	5
43	false	14.25	0.107502	5
44	false	12.125	0.063867	5
45	false	3.96875	0.09995	5
46	false	3.65625	0.035605	5
47	false	2.25	0.100725	5
48	false	41.5	0.091337	5
49	false	21.125	0.054426	5
50	false	10.125	0.097239	5
51	false	1.617188	0.028696	5
52	false	9.3125	0.056844	5
53	false	3.9375	0.03227	5
54	false	9.3125	0.082174	5
55	false	7.40625	0.033835	5
56	false	8.5625	0.042586	5
57	false	5.875	0.025338	5
58	false	7.4375	0.024443	5
59	false	2.609375	0.024383	5
60	false	2.46875	0.026061	5
61	false	2.265625	0.031289	5
62	false	3.796875	0.080472	5
63	false	5.40625	0.125151	5
64	false	8.5	0.05633	5
65	false	2.734375	0.09129	5
66	false	2.328125	0.030162	5
67	false	5.53125	0.028014	5
68	false	3.453125	0.045671	5
69	false	10.5625	0.031279	5
70	false	10.5625	0.07961	5
71	false	5.8125	0.146466	5
72	false	15.75	0.034988	5
73	false	17	0.047692	5
74	false	1.1875	0.022055	5
75	false	6.875	0.027778	5
76	false	6.1875	0.185093	5
77	false	20	0.366239	5
78	false	4.125	0.051786	5
79	false	4.4375	0.030682	5
80	false	21.5	0.098674	5
81	false	4.5	0.155597	5
82	false	7.0625	0.031177	5
83	false	5.5	0.053958	5
84	false	2.59375	0.030682	5
85	false	7.03125	0.177083	5
86	false	5	0.027143	5
87	false	2.875	0.032411	5
88	false	0.351563	0.02381	5
89	false	12.5625	0.126812	5
90	false	3.109375	0.025301	5
91	false	5.5625	0.029443	5
92	false	2.328125	0.05416	5
93	false	4.71875	0.022619	5
94	false	14.1875	0.05369	5
95	false	11.875	0.043211	5
96	false	3.765625	0.073602	5
97	false	3.15625	0.08546	5
98	false	3.046875	0.064286	5
99	false	10.5	0.059441	5

End of preview. Expand in Data Studio

Gemma-2-2B Toolformer: fused transcoder feature activations

Top-activating text snippets for every feature of the fused (base + delta) transcoders damaoo/gemma2b-toolformer-fused-transcoders (fused_delta_32_pctl_0.5), run on the fine-tuned model damaoo/gemma2b-toolformer. They were collected to power feature dashboards in a self-hosted Neuronpedia instance and exported from its Postgres database. Project code and write-up: github.com/damao2/Gemma-2-2B-Toolformer (phase2_neuronpedia/).

Feature indexing

Each of the 26 layers has a transcoder with 16384 Gemma-Scope base features (google/gemma-scope-2b-pt-transcoders, width 16k) plus 32 learned delta features appended at indices 16384–16415. is_delta = feature >= 16384.

Files

Config	Rows	Description
`features`	426,779	One row per feature that ever activated: `layer`, `feature`, `is_delta`, `max_act_approx`, `frac_nonzero`, `n_examples_general_8k`
`general_corpus_8k`	2,131,655	Top-5 snippets per feature over the first 8000 lines of a shuffled 20k corpus: 15k synthetic arithmetic / tool-call-intent prompts + 5k web snippets from C4
`toolcall_corpus_400`	2,109,261	Top-5 snippets per feature over the first 400 prompts of a 500-prompt corpus of tool-call positives, negatives and completions

Activation rows contain layer, feature, is_delta, max_value, max_token_index, tokens (list of token strings), values (per-token activation), and created_at.

Caveats

Snapshots, not complete runs. The collector flushed a cumulative top-k table to the DB periodically.
max_act_approx and frac_nonzero in features were written by the last flush (toolcall_corpus_400). frac_nonzero is computed over the stored top-k snippets only, so it is not the feature's true firing density.
Activations are on the raw transcoder scale. Delta features can reach values in the thousands.
The C4 portion of the snippets is distributed under C4's ODC-BY license; please attribute AllenAI's C4.
Neuronpedia's DB schema has many more per-feature columns (logit histograms, UMAP, etc.). The local instance never filled them, so they are omitted.

Example

from datasets import load_dataset
acts = load_dataset("damaoo/gemma2b-toolformer-feature-activations", "general_corpus_8k", split="train")
f = acts.filter(lambda r: r["layer"] == 20 and r["feature"] == 16399)   # a delta feature
for r in f: print(r["max_value"], "".join(r["tokens"]))

Downloads last month: 62