This dataset contains adversarial passages optimized to hijack retrieval for specific concepts. Access is granted for security research and evaluation.

TROPT Optimizer Benchmark — Corpus-Poisoning (Embedding) Triggers

Optimized adversarial passages from ranking 14 discrete optimizers on concept-specific corpus poisoning, plus their retrieval evaluation against MatanBT/msmarco-concepts.

Default FLOP budget per model

Every optimizer within a model runs under the same budget. Rows at this budget carry budget_label = "full".

model	encoder	params	default budget (FLOPs)
`minilm`	sentence-transformers/all-MiniLM-L6-v2	0.022B	3.6e14
`e5`	intfloat/e5-base-v2	0.110B	1.3e16
`qwen3emb`	Qwen/Qwen3-Embedding-0.6B	0.600B	1.3e16
`qwen3emb8b`	Qwen/Qwen3-Embedding-8B	7.570B	2.1e17

Grid

4 models x 14 optimizers x 8 concepts x 3 seeds x 2 trigger lengths (30, 100) = 2,688 runs, 26,880 rows.

gcg, mac, mcpal, gaslite, hotflip, autoprompt, arca, gbda, pal, ral, qcg, random_search, beast, adv_decoding.

Budget ladder

Each run appears at 10 budget points — snapshots of the same run truncated at a share of its budget. budget_flops is the absolute cap.

1pct, 2pct, 5pct, 10pct, 20pct, 30pct, 42pct, 60pct, 80pct, full.

Files

file	one row per
`triggers.parquet`	(run, `budget_label`)
`eval_triggered_messages/<model>.parquet`	trigger

Key columns

triggers.parquet — uid, model_short, optimizer_name, concept, seed, trigger_len, budget_label, budget_flops, best_trigger_str, best_loss, best_cos_sim, trigger_flops, n_steps_within_budget, optimized_instruction (template, containing {{OPTIMIZED_TRIGGER}}), adv_passage (template with the trigger substituted), mal_info, heldin_queries, run_id.

eval_triggered_messages/<model>.parquet — trigger_uid, heldin_cos_sim, heldout_cos_sim, heldin_mean_cos_sim, heldout_mean_cos_sim, heldin_appeared@10, heldout_appeared@10, heldin_ranks, heldout_ranks, n_corpus.

heldin_cos_sim / heldout_cos_sim are computed exactly as the training loss: cosine similarity to the centroid of that query split. *_ranks hold the per-query rank within the concept's corpus pool, so any k is recomputable.

Use `heldin_cos_sim`, not `best_cos_sim`, for realized performance

best_cos_sim is what the optimizer reached while working in token space. best_trigger_str is that token sequence decoded to text, and re-encoding text does not reliably reproduce the original tokens.

heldin_cos_sim is measured by re-embedding the published adv_passage, so it is what you actually get by planting the released text:

model	tokenizer	mean `heldin_cos_sim - best_cos_sim`	rows where the text is worse
`e5`	WordPiece	-0.0003	41%
`minilm`	WordPiece	-0.0010	40%
`qwen3emb`	BPE	-0.0086	92%
`qwen3emb8b`	BPE	-0.0172	94%

This tracks the tokenizer family, not the model. The two WordPiece encoders lose nothing systematic — ~40% of rows worse is symmetric float noise. Both BPE encoders lose one-directionally on >90% of rows, and the 8B loses twice what the 0.6B does. Adversarial token sequences do not survive a decode/encode round trip through BPE: even the token count is preserved in only ~65% of qwen3emb triggers.

Worst affected are the optimizers that select the most unusual tokens — on qwen3emb8b, gbda loses -0.098 (continuous relaxation lands between token boundaries), autoprompt -0.030 and arca -0.030.

The optimizer ranking is unaffected — it is computed from best_loss, uniformly across optimizers within a model.

Evaluation setting

Each concept's queries are ranked against that concept's own corpus split of MatanBT/msmarco-concepts (548–32,588 passages), not a web-scale index. appeared@k is therefore a rank within the concept pool; n_corpus is on every eval row, and pool sizes differ 59x across concepts.

Companion

MatanBT/tropt-optbench-triggers — the same benchmark in the LLM-jailbreak domain.

Downloads last month: 30

Total file size:

14.8 MB

You need to agree to share your contact information to access this dataset