Update dataset card to v3

README.md

@@ -52,10 +52,30 @@ A multi-label dataset mapping **CVE vulnerability descriptions** to their **CWE
 ## TL;DR
 
 - **Task:** given a CVE description, predict the CWE ID(s) — multi-label.
-- **69,386 examples** (capped/balanced build) — train 55,810 / validation 6,774 / test 6,802.
+- **71,640 examples** — train 50,074 / validation 11,052 / test 10,514. *(v3 — see "What's new in v3")*
 - **Format:** conversational `messages` (system / user / assistant), ready for chat-template SFT.
 - **Trust model:** labels are the **intersection** of two independent official sources (NVD + CNA), so single-source errors are excluded by construction.
 
+## What's new in v3
+
+This `v3` revision is re-weighted to train models on the cases they actually fail
+(inferring the CWE when the description does not name the weakness), and to evaluate
+them honestly. Changes vs. the previous revision:
+
+- **Low-information filter:** bare impact-title descriptions ("`<Product> <Impact>
+  Vulnerability`") that name no mechanism are dropped (~1.5% of rows; the CWE is
+  not derivable from the text).
+- **Hard-weighted training cap, train split only:** when a common CWE exceeds the
+  per-class cap (lowered to **1,500**), the **train** split keeps inference-"hard"
+  rows (weakness not named) over easy lexical-match rows. **Validation/test are
+  capped without this weighting**, so they keep real-world prevalence — train is
+  ~20% lexical-easy, val/test ~44%. Evaluate on the stratified easy/hard numbers.
+- **69 reviewed label corrections:** gold labels that two independent models flagged
+  and a human confirmed against the MITRE CWE definition (e.g. a CSRF labeled XSS,
+  an SSRF labeled "improper encoding") override the computed consensus.
+
+Models trained on the previous revision were trained on the un-reweighted data.
+
 ## Example row
 
 ```json
@@ -79,7 +99,6 @@ Multi-label targets are comma-separated and numerically sorted, e.g. `"CWE-79, C
 4. **Filtering:** REJECTED/DISPUTED CVEs, non-English and too-short (<40 char) descriptions are dropped; pseudo-labels (`NVD-CWE-noinfo`/`-Other`) are excluded.
 5. **De-duplication:** near-duplicate descriptions (MinHash, Jaccard ≥ 0.7) are grouped into families, and **whole families stay within a single split**, so near-identical text cannot leak across train/validation/test (verified: 0 shared descriptions across splits).
 6. **Split:** dedup-family-aware, ~80/10/10.
-7. **Balance (cap):** each CWE label is capped at **2,500 examples** (single seeded shuffle) so common classes don't swamp the long tail — see Statistics.
 
 ## Provenance (reproducibility)
 
@@ -91,8 +110,8 @@ Multi-label targets are comma-separated and numerically sorted, e.g. `"CWE-79, C
 
 ## Statistics
 
-- **Consensus keep rate:** 116,793 of 354,163 scanned CVEs reach consensus (~33%); most drops are "no consensus" (sources disagreed, or only one assigned a CWE).
-- **Balanced/capped build:** majority CWEs are then capped at **2,500 examples** each, reducing the 116,793 consensus rows to the **69,386 published rows** (train 55,810 / val 6,774 / test 6,802). This drops the top-10 label share from ~55% to **~31.5%**; all 117 labels appear in every split. The pre-cap natural distribution remains in this repo's git history (and is reproducible with `max_examples_per_cwe: null`).
+- **Keep rate:** 115,636 of 354,163 scanned CVEs reach consensus (~33%); after the low-information filter, the rarest-CWE floor, and hard-weighted capping, **71,640** rows are exported. Most drops are "no consensus" (sources disagreed, or only one assigned a CWE).
+- **Easy/hard composition (v3):** the weakness is lexically named in ~**20%** of train rows (hard-weighted) vs ~**44%** of validation/test (natural prevalence).
 - **Recency skew:** consensus concentrates on **2023+** CVEs, because that is when CNAs began reliably populating CWE — this is the main coverage bias.
 - **Label space:** 127 distinct View-1003 CWEs appear in the consensus set; a minimum-examples floor (50) prunes the rarest, leaving ~117 effective labels. The distribution is long-tailed (CWE-79 most common).
 - **Cardinality:** most examples carry a single label; ~14% carry two or more.