Upload README.md with huggingface_hub

README.md

@@ -1,172 +1,89 @@
 ---
-dataset_info:
-- config_name: access_events
-  features:
-  - name: at
-    dtype: float64
-  - name: pid
-    dtype: string
-  - name: token
-    dtype: string
-  - name: action
-    dtype: string
-  - name: reason
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 3987
-    num_examples: 72
-  download_size: 3738
-  dataset_size: 3987
-- config_name: chunks
-  features:
-  - name: chunk_id
-    dtype: string
-  - name: doc_id
-    dtype: string
-  - name: ordinal
-    dtype: int64
-  - name: text
-    dtype: string
-  - name: canary
-    dtype: string
-  - name: sensitivity
-    dtype: int64
-  - name: allow_tokens
-    list: string
-  - name: min_level
-    dtype: int64
-  - name: not_before
-    dtype: float64
-  - name: not_after
-    dtype: float64
-  splits:
-  - name: train
-    num_bytes: 6634093
-    num_examples: 14067
-  download_size: 1546333
-  dataset_size: 6634093
-- config_name: documents
-  features:
-  - name: doc_id
-    dtype: string
-  - name: title
-    dtype: string
-  - name: text
-    dtype: string
-  - name: department
-    dtype: string
-  - name: project
-    dtype: string
-  - name: region
-    dtype: string
-  - name: sensitivity
-    dtype: int64
-  - name: created_at
-    dtype: float64
-  - name: tags
-    list: string
-  - name: allow_tokens
-    list: string
-  - name: min_level
-    dtype: int64
-  - name: not_before
-    dtype: float64
-  - name: not_after
-    dtype: float64
-  splits:
-  - name: train
-    num_bytes: 5024896
-    num_examples: 4000
-  download_size: 1202298
-  dataset_size: 5024896
-- config_name: principals
-  features:
-  - name: pid
-    dtype: string
-  - name: name
-    dtype: string
-  - name: department
-    dtype: string
-  - name: title
-    dtype: string
-  - name: level
-    dtype: int64
-  - name: region
-    dtype: string
-  - name: projects
-    list: string
-  - name: tokens_at_t0
-    list: string
-  splits:
-  - name: train
-    num_bytes: 106613
-    num_examples: 600
-  download_size: 76132
-  dataset_size: 106613
-- config_name: queries
-  features:
-  - name: qid
-    dtype: string
-  - name: text
-    dtype: string
-  - name: asker
-    dtype: string
-  - name: kind
-    dtype: string
-  - name: topic
-    dtype: int64
-  - name: probe_target
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 32108
-    num_examples: 240
-  download_size: 9624
-  dataset_size: 32108
-- config_name: revocation_cases
-  features:
-  - name: pid
-    dtype: string
-  - name: token
-    dtype: string
-  - name: query_text
-    dtype: string
-  - name: target_chunk
-    dtype: string
-  - name: target_doc
-    dtype: string
-  - name: n_chunks_lost
-    dtype: int64
-  splits:
-  - name: train
-    num_bytes: 7356
-    num_examples: 50
-  download_size: 5064
-  dataset_size: 7356
+license: apache-2.0
+task_categories:
+- text-retrieval
+language:
+- en
+tags:
+- synthetic
+- rag
+- access-control
+- information-flow
+- benchmark
+- vector-search
+pretty_name: Clearance Benchmark (synthetic enterprise corpus with overlapping ACLs)
+size_categories:
+- 10K<n<100K
 configs:
-- config_name: access_events
-  data_files:
-  - split: train
-    path: access_events/train-*
-- config_name: chunks
-  data_files:
-  - split: train
-    path: chunks/train-*
 - config_name: documents
-  data_files:
-  - split: train
-    path: documents/train-*
+  data_files: documents/*.parquet
+- config_name: chunks
+  data_files: chunks/*.parquet
 - config_name: principals
-  data_files:
-  - split: train
-    path: principals/train-*
+  data_files: principals/*.parquet
+- config_name: access_events
+  data_files: access_events/*.parquet
 - config_name: queries
-  data_files:
-  - split: train
-    path: queries/train-*
+  data_files: queries/*.parquet
 - config_name: revocation_cases
-  data_files:
-  - split: train
-    path: revocation_cases/train-*
+  data_files: revocation_cases/*.parquet
 ---
+
+# Clearance Benchmark
+
+A **fully synthetic** enterprise corpus for measuring how well a retrieval system
+contains information flow: heavily overlapping ACLs, sensitivity levels,
+time-limited grants, a grant/revoke timeline, and evaluation queries.
+
+Generated by [`clearance.synth`](https://github.com/NagaYu/clearance) with seed `7`. **No real
+documents, no real access-control lists, and no real identities.** Regenerating
+with the same seed reproduces this dataset byte for byte.
+
+## Why this exists
+
+Query-time ACL filtering protects the index. It does not protect the summaries,
+agent memories and caches a RAG stack derives from that index. This dataset
+supports measuring three distinct leak channels — direct, existence-inference,
+and **derivative** — plus revocation lag, recall under high selectivity, and
+index memory.
+
+## Configs
+
+| Config | Rows | What it is |
+|---|---:|---|
+| `documents` | 4,000 | Documents with ACL (`allow_tokens`, `min_level`, embargo window) |
+| `chunks` | 14,067 | Indexed units; `canary` is a unique marker for exact leak detection |
+| `principals` | 600 | Departments, projects, clearance levels, capability tokens |
+| `access_events` | 72 | Grant/revoke timeline, including expiring secondments |
+| `queries` | 240 | Evaluation queries; `kind` is `normal` or `probe` |
+| `revocation_cases` | 50 | Triples where a revocation provably removes retrieved content |
+
+## Access semantics
+
+A principal may read a chunk when **all** of these hold:
+
+1. `allow_tokens ∩ principal_tokens ≠ ∅` (OR over capabilities),
+2. `principal.level >= min_level` (no read-up),
+3. `not_before <= t < not_after` (embargo window),
+4. the capability itself has not expired or been revoked at time `t`.
+
+## Canaries
+
+Every chunk carries a unique `CLR-<doc>-<ordinal>` marker inside its body text.
+Markers are stripped before embedding, so they never influence retrieval — they
+exist only so leak detection through derived text is exact rather than heuristic.
+
+## Intended use and limits
+
+Built to **measure defensive performance of index designs on synthetic data**.
+The `probe` queries target material the asking principal may not read; they
+exist to test whether a defence holds and are meaningful only because the
+generator knows the ground truth of a corpus it fabricated. This dataset is not
+an attack toolkit and carries no claim about any deployed system.
+
+Embedding-inversion attacks (recovering source text from vectors) are a real,
+documented risk for any shared vector store. They are referenced here as prior
+work and deliberately **not** implemented.
+
+Text is template-generated and is not a language-modelling benchmark. It exists
+to give retrieval something topically clustered to retrieve.