Adversarial triggers found by 15 discrete optimizers from TROPT against four embedding models, in GASLITE's concept-specific corpus-poisoning setting (Ben-Tov & Sharif, 2025): each trigger ends a passage carrying a malicious statement, so that the passage ranks high for one concept's MSMARCO queries. Each passage comes with its retrieval evaluation on the concept's held-in and held-out queries.

File	Rows	One row is
`triggers.parquet`	28,800	one optimization run's best trigger at one FLOP checkpoint
`eval_triggered_messages/<model_short>.parquet`	28,800	one trigger's retrieval evaluation on its concept's held-in and held-out queries

Models (4): sentence-transformers/all-MiniLM-L6-v2 (minilm), intfloat/e5-base-v2 (e5), Qwen/Qwen3-Embedding-0.6B (qwen3emb), Qwen/Qwen3-Embedding-8B (qwen3emb8b)
Optimizers (15): adv_decoding, arca, autoprompt, beast, gaslite, gbda, gcg, gcq, hotflip, mac, mcpal, pal, pez, ral, random_search. mcpal is PAL's search configuration (128 candidates, top-256 sampling, single-token replacement) with MAC's gradient momentum (μ = 0.6).
Loss: Similarity (minus the cosine similarity to the target) in every run.
Concepts (8): boston, flower, golf, iphone, mortgage, potter, sandwich, vaccine. Each passage is optimized on half of its concept's queries and evaluated on both halves.
Trigger lengths (2): 30 and 100 tokens.
Seeds (3): 42, 123, 777.

Loading

from datasets import load_dataset

R = "MatanBT/tropt-emb-optbench"
triggers = load_dataset(R, "triggers", split="triggers")
evals    = load_dataset(R, "eval_triggered_messages",
                        split="eval_triggered_messages")

Concepts

concept names a split of MatanBT/msmarco-concepts: its queries config holds the concept's MSMARCO queries, and its corpus config the n_corpus passages they are ranked against. Each run splits the concept's queries in half by its seed (random.Random(seed).shuffle of the queries in dataset order): the first half is held-in (heldin_queries) and sets the target, the second is held-out and is only evaluated on.

`triggers`

Each model has one FLOP budget per run, counting any auxiliary model the optimizer runs (google/gemma-2-2b-it, for beast and adv_decoding):

`model_short`	Budget (FLOPs)
`minilm`	`2.16e14`
`e5`	`7.8e15`
`qwen3emb`	`7.8e15`
`qwen3emb8b`	`1.26e17`

Every run is capped at 167% of its budget; hotflip and pez can stop earlier at 1,500 and 60,000 steps, and beast and adv_decoding after one step per trigger token. Its best trigger so far (lowest best_loss) is recorded at ten checkpoints, labeled by their share of the budget: 2pct, 3pct, 8pct, 17pct, 33pct, 50pct, 70pct, 100pct, 133pct, 167pct. A checkpoint excludes the step that crosses it, so trigger_flops never exceeds budget_flops; a checkpoint past a run's end holds the run's best trigger.

Identity

uid — {model_short}|{optimizer_name}|{concept}|s{seed}|t{trigger_len}|{budget_label}, e.g. qwen3emb|gcg|potter|s42|t30|100pct. Referenced by eval_triggered_messages.trigger_uid.
model_name, model_short — the target encoder.
optimizer_name — the optimization method.
loss_name — the training loss (Similarity).
concept — the concept attacked.
seed — the run's random seed; it also splits the queries.
trigger_len — the trigger's length in tokens (30 or 100).
run_id, run_name — wandb run identifiers, for provenance only.

Budget

budget_label, budget_fraction, budget_flops — the checkpoint: its label, its share of the model's budget, and its FLOPs.

Trigger & result

best_trigger_str — the optimized trigger.
best_loss — its training loss (lower is better).
best_cos_sim — -best_loss: the passage's cosine similarity to the target.
trigger_flops — cumulative FLOPs at the step that found this trigger.
n_steps_within_budget — optimizer steps counted toward this checkpoint.
run_total_flops, run_total_steps — the full run's totals.

Inputs

optimized_instruction — the passage template optimized: passage_prefix + mal_info + {{OPTIMIZED_TRIGGER}}, the placeholder the trigger replaces.
adv_passage — the template with best_trigger_str in place: the adversarial passage.
mal_info — the concept's malicious statement, verbatim from GASLITE.
passage_prefix, query_prefix — the encoder's input prefixes: passage: and query: for e5, an instruction before each query for the Qwen3 models, none for minilm.
initial_trigger — the trigger the run started from.
target_kind, heldin_queries, n_held_in, n_held_out — the target (heldin_query_centroid): the normalized mean embedding of the n_held_in held-in queries, each with query_prefix; n_held_out counts the held-out ones.

`eval_triggered_messages`

One file per model, with one row per triggers row (every checkpoint). The trigger's own encoder embeds adv_passage, the concept's queries (with query_prefix) and its corpus passages (with passage_prefix); for each query, the passage is ranked against the concept's corpus.

Trigger identity (copied from triggers)

trigger_uid — triggers.uid.
model_name, model_short, optimizer_name, loss_name, concept, seed, trigger_len, budget_label, best_cos_sim — as in triggers.

Retrieval ({split} is heldin or heldout)

{split}_cos_sim — the passage's cosine similarity to the centroid of the split's queries; heldin_cos_sim is the training objective, recomputed from adv_passage.
{split}_mean_cos_sim — the passage's mean cosine similarity to the split's queries.
{split}_appeared@10 — the share of the split's queries whose top 10 results include the passage.
{split}_ranks — the passage's rank against the concept's corpus, per query (1 is first; 101 means below the top 100); heldin_ranks is aligned with triggers.heldin_queries.
n_{split}_eval — the number of queries in the split.
n_corpus — the number of passages in the concept's corpus.

For example, the mean held-out appeared@10 per optimizer and model, for 30-token triggers at the budget (100pct):

df = evals.to_pandas()
heldout = (df[(df.trigger_len == 30) & (df.budget_label == "100pct")]
             .groupby(["model_short", "optimizer_name"])["heldout_appeared@10"]
             .mean().unstack(0))