Update v2 README to refreshed splits + directgen layout and counts

v2/README_v2_release.md

@@ -1,36 +1,58 @@
-# V2 Release (Polymarket Trading-Layer, pre80m)
-
-This V2 release contains the finalized trading-layer tables used for modeling.
-
-## Files
-
-1. `v2/v2_behavior_features_merged_pospnl_kline_v1.parquet`
-   - Final modeling table (behavior + position/pnl features + 12h/2h Kline features).
-
-2. `v2/v2_behavior_main_pre80m_normal_final_cap30_pricecovered.parquet`
-   - Final behavior base table after filtering (normal users, cap30, price-covered markets).
-
-3. `v2/v2_market_semantic_from_behavior_cap30_pricecovered.parquet`
-   - Market semantic lookup aligned with the final behavior table.
-
-4. `v2/v2_trade_price_matched_pre80m_cap30.parquet`
-   - Matched trade-price table restricted to the final market universe.
-
-## Scope
-
-- Time cutoff: pre80m
-- User-market cap: 30
-- Markets included: only price-covered markets
-
-## Kline window semantics (2-hour bins)
-
-For the final modeling table (`v2_behavior_features_merged_pospnl_kline_v1.parquet`), Kline features are computed in rolling pre-trade windows with 2-hour width:
-
-- `k0_*`: 0–2 hours before trade
-- `k1_*`: 2–4 hours before trade
-- `k2_*`: 4–6 hours before trade
-- `k3_*`: 6–8 hours before trade
-- `k4_*`: 8–10 hours before trade
-- `k5_*`: 10–12 hours before trade
-
-Each window includes: `open`, `high`, `low`, `close`, `points`, `avg`.
+# V2 Release (Trading-Level, refreshed)
+
+This V2 release is organized into three subfolders under `v2/`:
+
+- `v2/source/` — source parquet tables
+- `v2/splits/` — structured train/validation/test splits (parquet)
+- `v2/benchmark_directgen/` — DirectGen benchmark (jsonl.gz)
+
+## 1) Source tables (`v2/source`)
+
+- `behavior_fixcost.parquet`
+- `trade_price_matched.parquet`
+- `market_semantic.parquet`
+- `source_metadata.json`
+
+## 2) Dataset splits (`v2/splits`)
+
+Files:
+- `train.parquet`
+- `validation.parquet`
+- `test.parquet`
+- `split_metadata.json`
+- `wallet_ids_experiment_2000.txt`
+
+Current split policy:
+- Fixed wallet universe: 2000 wallets
+- Per-wallet temporal split, then global merge
+- Current ratio: **85/10/5**
+- `validation/test` enforce **prior_same_market_trades >= 1** on final retained rows
+
+Current row counts:
+- train: **1,311,941**
+- validation: **114,471**
+- test: **59,560**
+
+## 3) DirectGen benchmark (`v2/benchmark_directgen`)
+
+Files:
+- `train.jsonl.gz`
+- `validation.jsonl.gz`
+- `test.jsonl.gz`
+- `benchmark_metadata.json`
+
+Current generation policy:
+- History organized as market blocks (`HistoryJSON`)
+- Other-market context: top-5 most recently active markets
+- Strict causality (`history < current sample time`)
+- History amount field: `amount_usdc`
+- Target amount field: `amount_usdc_est`
+- History action labels unified to target direction vocabulary:
+  - `increase`, `decrease`, `close`
+
+Current row counts:
+- train: **1,311,941**
+- validation: **114,471**
+- test: **59,559**
+
+> Note: benchmark test can be 1 row fewer than split test due to strict empty-history filtering during benchmark generation.