Upload README.md with huggingface_hub

README.md

@@ -49,50 +49,36 @@ configs:
   data_files:
   - split: full
     path: hackersignal_sample_10k.jsonl
-- config_name: task1_exploit_clf
+- config_name: task1_cve_linkage
   data_files:
   - split: train
-    path: benchmark/task1_exploit_clf/train.jsonl
+    path: benchmark_v2/task1_cve_linkage/train.jsonl
   - split: validation
-    path: benchmark/task1_exploit_clf/val.jsonl
+    path: benchmark_v2/task1_cve_linkage/val.jsonl
   - split: test
-    path: benchmark/task1_exploit_clf/test.jsonl
-- config_name: task2_cve_linkage
-  data_files:
-  - split: train
-    path: benchmark/task2_cve_linkage/train.jsonl
-  - split: validation
-    path: benchmark/task2_cve_linkage/val.jsonl
-  - split: test
-    path: benchmark/task2_cve_linkage/test.jsonl
+    path: benchmark_v2/task1_cve_linkage/test.jsonl
   - split: corpus
-    path: benchmark/task2_cve_linkage/corpus.jsonl
-- config_name: task3_severity
-  data_files:
-  - split: train
-    path: benchmark/task3_severity/train.jsonl
-  - split: validation
-    path: benchmark/task3_severity/val.jsonl
-  - split: test
-    path: benchmark/task3_severity/test.jsonl
-- config_name: task4_hacker_exploit_labeling
+    path: benchmark_v2/task1_cve_linkage/corpus.jsonl
+- config_name: task2_signal_detection
   data_files:
   - split: train
-    path: benchmark/task4_hacker_exploit_labeling/train.jsonl
+    path: benchmark_v2/task2_signal_detection/train.jsonl
   - split: validation
-    path: benchmark/task4_hacker_exploit_labeling/val.jsonl
+    path: benchmark_v2/task2_signal_detection/val.jsonl
   - split: test
-    path: benchmark/task4_hacker_exploit_labeling/test.jsonl
-- config_name: task5_hacker_signal_detection
+    path: benchmark_v2/task2_signal_detection/test.jsonl
+  - split: corpus
+    path: benchmark_v2/task2_signal_detection/corpus.jsonl
+- config_name: task3_temporal_generalization
   data_files:
   - split: train
-    path: benchmark/task5_hacker_signal_detection/train.jsonl
+    path: benchmark_v2/task3_temporal_generalization/train.jsonl
   - split: validation
-    path: benchmark/task5_hacker_signal_detection/val.jsonl
+    path: benchmark_v2/task3_temporal_generalization/val.jsonl
   - split: test
-    path: benchmark/task5_hacker_signal_detection/test.jsonl
+    path: benchmark_v2/task3_temporal_generalization/test.jsonl
   - split: corpus
-    path: benchmark/task5_hacker_signal_detection/corpus.jsonl
+    path: benchmark_v2/task3_temporal_generalization/corpus.jsonl
 ---
 
 # HackerSignal
@@ -108,7 +94,7 @@ A large-scale, multi-source dataset linking hacker community discourse, exploit
 | Source layers | 8 |
 | Temporal span | 1988--2026 |
 | CVE-linked rows | 360,004 |
-| Benchmark tasks | 5 |
+| Benchmark tasks | 3 |
 
 ## Quick Start
 
@@ -119,7 +105,7 @@ from datasets import load_dataset
 sample = load_dataset("DatasetSubmission/HackerSignal", "sample")
 
 # Load a benchmark task
-task1 = load_dataset("DatasetSubmission/HackerSignal", "task1_exploit_clf")
+task1 = load_dataset("DatasetSubmission/HackerSignal", "task1_cve_linkage")
 
 # Load the full corpus (12 GB)
 full = load_dataset("DatasetSubmission/HackerSignal", "default")
@@ -140,17 +126,8 @@ full = load_dataset("DatasetSubmission/HackerSignal", "default")
 
 ## Benchmark Tasks
 
-### Task 1: Exploit Relevance Classification
-Binary classification (exploit-relevant vs. not). Cross-source labeling: positives are CVE-mentioning posts from exploit/advisory sources; negatives are non-CVE hacker community posts. Temporal split: train (<2022), val (2022--2023), test (2024+).
-
-| Split | Rows |
-|-------|------|
-| Train | 89,906 |
-| Val | 1,400 |
-| Test | 1,236 |
-
-### Task 2: CVE Linkage Retrieval
-Given source evidence text, retrieve the correct NVD CVE entry from a corpus of 340K descriptions. Quality-controlled to remove unretrievable rows.
+### Task 1: CVE Linkage Retrieval (CVE-R)
+Cross-source temporally OOD entity grounding: given exploit/advisory evidence text, retrieve the correct NVD CVE entry from a corpus of 340K descriptions. Quality-controlled to remove unretrievable rows.
 
 | Split | Rows |
 |-------|------|
@@ -159,32 +136,24 @@ Given source evidence text, retrieve the correct NVD CVE entry from a corpus of
 | Test | 550 |
 | Corpus | 340,536 |
 
-### Task 3: Severity Prediction
-4-class CVSS severity prediction (low/medium/high/critical) from advisory text with explicit severity strings stripped.
-
-| Split | Rows |
-|-------|------|
-| Train | 6,797 |
-| Val | 12,836 |
-| Test | 11,711 |
-
-### Task 4: Hacker Exploit Labeling
-Ternary weak-supervision task: 0=non-exploit/noise, 1=vulnerability discussion, 2=actionable exploit.
+### 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.
 
 | Split | Rows |
 |-------|------|
-| Train | 200,382 |
-| Val | 1,933 |
-| Test | 100,963 |
+| Train | 4,202 |
+| Val | 1,606 |
+| Test | 407 |
+| Corpus | 340,536 |
 
-### Task 5: Hacker Exploit Signal Detection
-Joint detection + retrieval: identify actionable exploit signals and link to NVD CVE context. 10:1 negative-to-positive ratio reflects real-world signal rarity.
+### 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.
 
 | Split | Rows |
 |-------|------|
-| Train | 4,202 |
-| Val | 1,606 |
-| Test | 407 |
+| Train | 41,630 |
+| Val | 1,776 |
+| Test | 393 |
 | Corpus | 340,536 |
 
 ## Schema