Update Task 2 config to exploit_type; add known-issues section (viewer count, unsafe-scan note)

README.md

@@ -42,16 +42,14 @@ configs:
     path: benchmark_v2/task1_cve_linkage/test.jsonl
   - split: corpus
     path: benchmark_v2/task1_cve_linkage/corpus.jsonl
-- config_name: task2_signal_detection
+- config_name: task2_exploit_type
   data_files:
   - split: train
-    path: benchmark_v2/task2_signal_detection/train.jsonl
+    path: benchmark_v2/task2_exploit_type/train.jsonl
   - split: validation
-    path: benchmark_v2/task2_signal_detection/val.jsonl
+    path: benchmark_v2/task2_exploit_type/val.jsonl
   - split: test
-    path: benchmark_v2/task2_signal_detection/test.jsonl
-  - split: corpus
-    path: benchmark_v2/task2_signal_detection/corpus.jsonl
+    path: benchmark_v2/task2_exploit_type/test.jsonl
 - config_name: task3_temporal_generalization
   data_files:
   - split: train
@@ -119,15 +117,14 @@ Cross-source temporally OOD entity grounding: given exploit/advisory evidence te
 | Test | 550 |
 | Corpus | 340,536 |
 
-### Task 2: Hacker Signal Detection (HSD)
-Two-stage task: (1) detect actionable exploit signals from hacker community noise, (2) ground positives to NVD CVE. 10:1 negative-to-positive ratio reflects real-world signal rarity.
+### Task 2: Exploit Type Classification (ETC)
+8-class temporal OOD classification of exploit posts by type: injection, XSS, memory corruption, DoS, file inclusion, authentication/access bypass, RCE, and information disclosure. Sourced from ExploitDB and HackerOne public reports; split by publication year so test posts postdate all training examples.
 
 | Split | Rows |
 |-------|------|
-| Train | 4,202 |
-| Val | 1,606 |
-| Test | 407 |
-| Corpus | 340,536 |
+| Train | 64,413 |
+| Val | 4,735 |
+| Test | 1,735 |
 
 ### Task 3: Temporal Generalization (TG)
 Identical retrieval formulation to Task 1 but with strict CVE-disjoint constraint: C_train ∩ C_test = ∅. Split by CVE publication year (train: <2022, test: 2024+) to test generalization to wholly unseen vulnerabilities.
@@ -172,6 +169,11 @@ This dataset is intended for **defensive cybersecurity research** only. Prohibit
 - Operational blocking or law-enforcement decisions based solely on model outputs
 - Republishing text from `metadata_or_pointer_only` sources
 
+## Known Issues
+
+- **HuggingFace dataset viewer row count**: The viewer may display a partial count (~1M) for the `default/full` split because HF's streaming viewer cannot fully index files over ~10 GB. The actual row count is 7,447,646 as validated against the SHA-256 manifest.
+- **Content-scanner warnings**: HF's automated scanner flags several files (including the sample JSONL) as "Unsafe" due to exploit and vulnerability text. This is expected for a cybersecurity research dataset and does not indicate malicious content. See the Responsible Use section below.
+
 ## Limitations
 
 - English-language bias; non-English communities may be underrepresented