JamPivot

A benchmark of typed anti-jamming decisions from spectrum waterfalls and text constraints

In a band shared with jammers and protected users, a radio link must keep choosing its channel and transmission mode. It sees the interferers in a spectrum waterfall, but the constraints that decide the best action (the required rate, the power budget, the switching cost and the users to spare) vary between deployments and do not appear in the waterfall. Each JamPivot scene is one such decision: a two-panel waterfall image, the constraints as text, and typed questions whose options are executable actions, labelled by the one-step Bayes oracle.

JamPivot is the benchmark of SpecOne; the training and inference code is at github.com/dan-zong-hao/SpecOne.

At a glance

• 48,000 scenes: 40,000 training scenes (the last 600 are a monitoring set) and 8,000 test scenes, plus the 15,994 scenes of 7,997 training contrast pairs.
• Eleven interferer categories: spot, comb, partial-band, barrage, pulsed, sweep, random hop, tracking, reactive, two jammers, and coexistence with protected users. Jamming scenes (70%) hold one jammer or, in a quarter of them, two; coexistence scenes (30%) hold one or two protected users, half of them with a jammer.
• Four typed questions per scene: chan (which of 9 channels), mode (normal, boost, spread4, spread16, mute, asked on a chosen channel), feas (is the link feasible) and the auxiliary type (which jammer).
• Oracle labels: every scene stores the expected utility of all 45 channel–mode actions over the next slot under the known interferer laws (random hops and Markov users averaged over 32 rollouts), from which the optimal sets, the one-hot targets and the utility-soft targets (τ = 0.05) are derived.

Split	Scenes	Question rows	Shards	Purpose
train	40,000	148,187	40	training (scenes 39,400–39,999: monitoring only)
pair_train	15,994	59,393	16	both scenes of 7,997 training contrast pairs
test_id	3,000	11,085	3	in-distribution test
ood_combo	1,000	4,000	1	three jammer pairs never seen in training
ood_style	1,000	3,711	1	a held-out phrasing of the constraints
pair_test	3,000	11,160	3	1,500 test contrast pairs: one edit of the text, the image or the role changes the best action

Files

data/<split>/<split>-NNNNN.tar   WebDataset shards, 1,000 scenes each: <scene>.png and <scene>.json
rows_<split>.jsonl               one row per question, in OneJev's row format
scenes_<split>.jsonl             one record per scene: generator spec, utility table, feasibility, jammer type
pairs_pair_train.jsonl, pairs_pair_test.jsonl   contrast pairs: pair id, edit kind, scene a, scene b
train_softpairs.jsonl, dev.jsonl the exact SpecOne training and monitoring files (specone/build_train.py)
build_summary.json               generator settings

In the shards, <scene>.png is the 400 × 530 waterfall and <scene>.json is {"scene": <scene record>, "rows": [<its question rows>]}, so the Hugging Face viewer and datasets can stream image and labels together.

Row and scene record fields

A row (rows_<split>.jsonl, also inside each <scene>.json):

Field	Content
id	<scene>-chan, -mode, -feas or -type
task	chan, mode, feas or type
state	the text state: waterfall layout, our link, link budget, rate requirement, power boost, channel switch cost, protected users; mode rows add chosen channel
question	instructions and the lettered options (criteria)
target	uniform distribution over the optimal options
target_soft	utility-soft distribution, τ = 0.05
media	path of the waterfall image, media/<split>/<scene>.png
meta	scene id, family (jam / coex), split, phrasing style, jammer type, optimal options; mode rows: the channel asked on

A scene record (scenes_<split>.jsonl):

Field	Content
spec	seed, family, our SNR, jammers (kind, J/S, parameters), protected users, switching cost, boost, rate floor, guard band
f_prev	our current channel (0-based)
utility	9 × 5 expected utility of every channel–mode action, null where the action is excluded
success, violation	expected share of successful milliseconds per action; probability of hitting a protected user per channel
feasible, jam_type	the feasibility label; the jammer type (multiple for two jammers, null for coexistence)

Usage

With the SpecOne code (it expects DATA/rows_<split>.jsonl and DATA/media/<split>/<scene>.png):

hf download danzonghao/JamPivot --repo-type dataset --local-dir JamPivot
cd JamPivot
for f in data/*/*.tar; do s=$(basename "$(dirname "$f")"); mkdir -p media/$s; tar -xf "$f" -C media/$s --wildcards '*.png'; done

With datasets (streams images and labels):

from datasets import load_dataset
ds = load_dataset("danzonghao/JamPivot", split="test_id", streaming=True)
ex = next(iter(ds))
ex["png"]          # PIL image of the waterfall
ex["json"]["rows"] # the scene's question rows

License

CC BY 4.0. All waterfalls are generated by the JamPivot simulator. The setting and four of the jammers follow Liu et al., "Anti-jamming communications using spectrum waterfall: a deep reinforcement learning approach", IEEE Commun. Lett., 2018; the reactive jammer follows Li et al., arXiv 2502.02385.

Citation

The paper is under review; the BibTeX entry will be added here once it is published.