Update README with dataset release details

README.md

@@ -26,6 +26,75 @@ It includes route-related information such as coordinates, station sequences, tr
 
 The current repository contains multiple CSV exports, including benchmark-specific train/test splits (`transit-lm-benchmark*-train.csv`, `transit-lm-benchmark*-test.csv`), general benchmark test splits (`transit-lm-general-benchmark*-test.csv`), and a static station information file (`transit-lm-station-info.csv`).
 
+The `transit-lm-benchmark*-train.csv` files are the benchmark-aligned SFT training splits. The `transit-lm-benchmark*-test.csv` files are the corresponding benchmark test splits used for task evaluation. The `transit-lm-general-benchmark*-test.csv` files are the real-world benchmark test sets intended for evaluating general-purpose LLMs on natural station names and coordinates. The `transit-lm-station-info.csv` file provides supporting station-level reference information for the transit network.
+
+## Dataset Release & Evaluation Protocol
+
+All released data follows the anonymization principles described in our paper.
+
+**Overview.** We release two parallel test sets that correspond one-to-one, serving different model categories:
+
+- **Anonymized SFT test set** (30K samples). For models trained on our anonymized corpus. These models have only seen numeric station IDs and perturbed coordinates, so they must be evaluated in the same representation space.
+- **Real-world benchmark** (30K samples). For general-purpose LLMs not trained on our data. These models operate with real station names and coordinates and cannot interpret anonymized IDs, so they require unperturbed evaluation data.
+
+This parallel design ensures fair evaluation for both model types while keeping results directly comparable.
+
+## 1. Benchmark Test Samples
+
+30,000 test examples across three tasks: optimal route generation, preference-aware planning, and multi-route generation.
+
+**Data characteristics:**
+- Real GPS coordinates without perturbation
+- Real station names in natural language
+- Full route-structural metadata: station sequences, line sequences, transfer points, distances, travel times, and fares
+
+**Purpose.** This subset is the ground-truth evaluation benchmark for the research community. The CPT/SFT training corpus applies coordinate perturbation, so evaluating general-purpose LLMs on perturbed data would not reflect real-world performance. Unperturbed samples enable fair comparison across models trained on our corpus, tool-augmented LLMs, and any future transit-domain model.
+
+**Privacy.** These samples contain no user identifiers, no timestamps, and no cross-record linkage. Each is an isolated OD request. All station coordinates and names are already publicly accessible through mapping APIs, so releasing them in benchmark format introduces no additional privacy risk.
+
+## 2. CPT + SFT Corpus (Anonymized)
+
+13.9M CPT records, 90K SFT training examples, and 30K SFT test examples.
+
+### Anonymization
+
+| Transformation | Method | Training Impact |
+|---|---|---|
+| Coordinate perturbation | Deterministic hash-based offset, +/-100m on a ~500m grid | Negligible. Grid-coherent offset preserves spatial structure. Error < 0.1% of mean route length |
+| Station name removal | Names removed from station sequences, retaining only numeric IDs | None. Topological learning relies on IDs and connectivity |
+
+**Rationale.** Bulk harvesting coordinate-ID mappings from a 13.9M-record corpus could enable unauthorized commercial use. OD coordinates reflect real user trip intentions. Perturbation combined with structural safeguards prevents re-identification.
+
+### SFT Test Set
+
+The 30K SFT test samples undergo identical anonymization. See the Overview above for how they relate to the real-world benchmark and which model category uses which test set.
+
+### Usability
+
+Despite anonymization, this corpus fully supports continual pre-training for transit topology and spatial reasoning, supervised fine-tuning for all three benchmark tasks, and reproducibility of all paper results.
+
+## 3. Privacy Architecture
+
+Unlike GPS trajectory datasets such as T-Drive or GeoLife that record continuous multi-day traces enabling re-identification, each record here is an isolated OD planning request with no temporal continuity.
+
+Structural safeguards enforced at the data construction stage:
+- Single calendar day sampling with no timestamps retained
+- User identifiers removed prior to construction, with no linkage key across records
+- Only route-structural metadata released. No demographic attributes, device fingerprints, or PII
+
+These properties make it infeasible to associate records with the same individual or reconstruct mobility patterns, regardless of coordinate perturbation.
+
+## 4. Why Both Perturbed and Real Data?
+
+| Concern | Anonymized corpus | Benchmark samples |
+|---|---|---|
+| Coordinate privacy | Yes. +/-100m perturbation prevents bulk harvesting | Public data, no new risk |
+| Fair evaluation | No. Perturbed coordinates disadvantage general LLMs | Yes. Real coordinates enable direct comparison |
+| Reproducibility | Yes. Full training pipeline reproducible | Yes. Standardized community test set |
+| Training utility | Yes. < 0.1% spatial error, topology preserved | Not used for training |
+
+This dual-release design maximizes research utility through reproducible training and fair evaluation, while minimizing privacy exposure by applying perturbation only where data scale creates risk.
+
 ## Dataset Status
 
 This repository is currently being uploaded incrementally. The files currently visible on Hugging Face may represent only part of the full dataset.