Datasets:

nac7
/

tool-output-faithfulness

id stringlengths 5 5	messages listlengths 14 30	probe stringclasses 20 values	required_facts listlengths 1 1	probe_type stringclasses 2 values	depth stringclasses 2 values	concept stringclasses 7 values	placement stringclasses 2 values
s0000	[ { "role": "system", "content": "You are a terse engineering assistant.", "name": null }, { "role": "user", "content": "Also tweak the test fixtures (pass 0); minor cleanup.", "name": null }, { "role": "assistant", "content": "Updated the test fixtures; no functional change.",...	What database backs the service again?	[ "PostgreSQL" ]	paraphrase	mid	For	statement
s0001	[ { "role": "system", "content": "You are a terse engineering assistant.", "name": null }, { "role": "user", "content": "Also tweak the CI config (pass 0); minor cleanup.", "name": null }, { "role": "assistant", "content": "Updated the CI config; no functional change.", "na...	What signing algorithm do our JWT tokens use?	[ "RS256" ]	verbatim	mid	Authentication	statement
s0002	[ { "role": "system", "content": "You are a terse engineering assistant.", "name": null }, { "role": "user", "content": "Also tweak the build script (pass 0); minor cleanup.", "name": null }, { "role": "assistant", "content": "Updated the build script; no functional change.", ...	How long do we keep entries before they expire?	[ "900 seconds" ]	paraphrase	mid	The	buried
s0003	[ { "role": "system", "content": "You are a terse engineering assistant.", "name": null }, { "role": "user", "content": "Set up the relevant component (step 0).", "name": null }, { "role": "assistant", "content": "We deploy to the Kubernetes namespace 'orion-prod'.", "name"...	Which Kubernetes namespace do we deploy to?	[ "orion-prod" ]	verbatim	early	We	statement
s0004	[ { "role": "system", "content": "You are a terse engineering assistant.", "name": null }, { "role": "user", "content": "Also tweak the dev container (pass 0); minor cleanup.", "name": null }, { "role": "assistant", "content": "Updated the dev container; no functional change.",...	What is the API rate limit per key?	[ "100 requests" ]	verbatim	mid	The	statement
s0005	[ { "role": "system", "content": "You are a terse engineering assistant.", "name": null }, { "role": "user", "content": "Set up the relevant component (step 0).", "name": null }, { "role": "assistant", "content": "Background jobs run through a Redis-backed queue.", "name": ...	Which system processes async tasks?	[ "[Rr]edis" ]	paraphrase	early	Background	statement
s0006	[ { "role": "system", "content": "You are a terse engineering assistant.", "name": null }, { "role": "user", "content": "Also tweak the makefile (pass 0); minor cleanup.", "name": null }, { "role": "assistant", "content": "Updated the makefile; no functional change.", "name...	How long before we give up waiting on a downstream call?	[ "30 seconds" ]	paraphrase	mid	The	buried
s0007	[ { "role": "system", "content": "You are a terse engineering assistant.", "name": null }, { "role": "user", "content": "Set up the relevant component (step 0).", "name": null }, { "role": "assistant", "content": "Production runs in the us-east-1 region.", "name": null },...	Where is the live environment hosted geographically?	[ "us-east-1" ]	paraphrase	early	Production	statement
s0008	[ { "role": "system", "content": "You are a terse engineering assistant.", "name": null }, { "role": "user", "content": "Also tweak the build script (pass 0); minor cleanup.", "name": null }, { "role": "assistant", "content": "Updated the build script; no functional change.", ...	Which framework is the service built on?	[ "[Ff]ast[Aa]pi" ]	verbatim	mid	The	statement
s0009	[ { "role": "system", "content": "You are a terse engineering assistant.", "name": null }, { "role": "user", "content": "Also tweak the dev container (pass 0); minor cleanup.", "name": null }, { "role": "assistant", "content": "Updated the dev container; no functional change.",...	Where are the secrets stored?	[ "Secrets Manager" ]	verbatim	mid	Secrets	statement
s0010	[ { "role": "system", "content": "You are a terse engineering assistant.", "name": null }, { "role": "user", "content": "Set up the relevant component (step 0).", "name": null }, { "role": "assistant", "content": "Tests are green and the lint passes. For the primary datastore w...	What database backs the service again?	[ "PostgreSQL" ]	paraphrase	early	For	buried
s0011	[ { "role": "system", "content": "You are a terse engineering assistant.", "name": null }, { "role": "user", "content": "Also tweak the dev container (pass 0); minor cleanup.", "name": null }, { "role": "assistant", "content": "Updated the dev container; no functional change.",...	How do we verify who a caller is?	[ "RS256" ]	paraphrase	mid	Authentication	buried
s0012	[ { "role": "system", "content": "You are a terse engineering assistant.", "name": null }, { "role": "user", "content": "Also tweak the docs (pass 0); minor cleanup.", "name": null }, { "role": "assistant", "content": "Updated the docs; no functional change.", "name": null ...	How long do we keep entries before they expire?	[ "900 seconds" ]	paraphrase	mid	The	statement
s0013	[ { "role": "system", "content": "You are a terse engineering assistant.", "name": null }, { "role": "user", "content": "Set up the relevant component (step 0).", "name": null }, { "role": "assistant", "content": "I went ahead and wired up the module. We deploy to the Kubernete...	Where does the service ship to in the cluster?	[ "orion-prod" ]	paraphrase	early	We	buried
s0014	[ { "role": "system", "content": "You are a terse engineering assistant.", "name": null }, { "role": "user", "content": "Also tweak the logging format (pass 0); minor cleanup.", "name": null }, { "role": "assistant", "content": "Updated the logging format; no functional change....	How many calls can a client make before it gets throttled?	[ "100 requests" ]	paraphrase	mid	The	statement
s0015	[ { "role": "system", "content": "You are a terse engineering assistant.", "name": null }, { "role": "user", "content": "Set up the relevant component (step 0).", "name": null }, { "role": "assistant", "content": "Background jobs run through a Redis-backed queue.", "name": ...	What backs our background job queue?	[ "[Rr]edis" ]	verbatim	early	Background	statement

End of preview. Expand in Data Studio

Agentic Context-Compression Faithfulness Benchmark (Tier A)

A reproducible, judge-free benchmark measuring whether context compression preserves the specific facts a later agent turn depends on — in multi-turn agentic sessions, not single-shot prompts.

📄 Paper / write-up: Faithful Context Compression for Agentic Sessions: A Reproducible Benchmark and a Granularity-vs-Signal Decomposition (preprint).
🛠️ Harness + generator (MIT): https://github.com/nac7/ctx-gate — one command reproduces every number.
🏆 Leaderboard: see LEADERBOARD.md in the repo.

What's in it

Synthetic agentic sessions (system prompt + tool dumps + re-read files + filler turns + task shifts) each carrying one planted fact (a realistic service-configuration value with a known ground-truth answer) and a probe that requires it. The primary Layer-A metric is a string match on the compressed context, so it is immune to pretraining contamination regardless of how common the fact is; Layer-B controls for prior knowledge via a compressed−full delta (see limitations).

Each item:

field	description
`id`	unique item identifier
`messages`	the full session up to (not including) the probe
`probe`	the question asked now
`probe_type`	`verbatim` (shares content words with the fact) or `paraphrase` (shares meaning, zero content words)
`required_facts`	list of ground-truth answer string(s), e.g. `["PostgreSQL"]` (⊆ the planted fact)
`depth` / `placement`	where the fact sits (`early`/`mid`/`recent`, `statement`/`buried`)
`concept`	concept tag derived from the fact

The paraphrase split is the discriminating stressor: methods that match on surface form pass verbatim and fail paraphrase.

How it's scored (two layers)

Layer A — fact retention (primary, deterministic, $0): does the compressed context still contain the planted fact (exact match)? Decoupled from any model and from any compressor's selection criterion.
Layer B — answer accuracy (optional): a reader model answers the probe under compressed vs. full context; report Δ. Confirms retention predicts accuracy.

Compare methods at matched realized savings, not matched knob.

Quickstart

from datasets import load_dataset
ds = load_dataset("nac7/tool-output-faithfulness", split="train")
print(ds[0]["probe"], ds[0]["probe_type"], ds[0]["required_facts"])

Reproduce the full benchmark (methods, seeds, Pareto):

pip install "ctx-gate[bench]"
ctx-gate bench scale --n 300 --seeds 0,1,2 \
  --methods ctxgate,ctxgate_sentence,ctxgate_sentence_embedding,selective \
  --out results/scale

Headline finding

Token-level self-information pruning (Selective Context) is perfectly lossless (100% retention, zero between-seed variance) through ~89% savings and Pareto-dominates message-level relevance selection. Decomposing the gap shows granularity (message→sentence: +63.6 pts paraphrase retention), not the selection signal, is the dominant factor — and the winning method is query-agnostic, so query-awareness is not the operative variable. Self-information is near-optimal here because planted facts are high-surprisal; the benchmark predicts this inverts for low-surprisal facts (the open frontier).

Intended uses & limitations

Use: benchmarking context/prompt compressors on faithfulness; ablating granularity vs. selection signal; as a fast, judge-free CI gate for compressors.
Limitation (important): facts are high-surprisal by construction — the regime most favorable to self-information pruning. Do not read the result as "self-information is universally optimal." A low-surprisal fact suite and a real-transcript (Tier B) tier are planned for external validity.
Contamination: Layer-A is contamination-immune by construction (string match on the context, not a model guess). Some Layer-B answers are common technical terms; the compressed−full delta controls for this. Regenerate with a fresh seed / canary (or fictional-constant facts) to further harden Layer-B.

Datasheet & license

Follows Datasheet for Datasets (Gebru et al., 2021) — see the repo's paper/dataset-release-plan.md. Data: CC-BY-4.0. Generator/harness code: MIT. DOI: 10.57967/hf/9671 (pinned version v1.0).

Citation

@misc{lele2026faithfulcompression,
  title  = {Faithful Context Compression for Agentic Sessions:
            A Reproducible Benchmark and a Granularity-vs-Signal Decomposition},
  author = {Lele, Nachiket},
  year   = {2026},
  doi    = {10.57967/hf/9671},
  note   = {Benchmark: huggingface.co/datasets/nac7/tool-output-faithfulness (v1.0);
            code: github.com/nac7/ctx-gate},
}

Downloads last month: 81