Add 3x-context dataset variant (trainset --radius)

- build_trainset gains --radius (chars each side of the cited name) and --out;
  merge-gap scales with radius. Default 600 unchanged.
- trainset_3x + train/val/test_3x.jsonl: same 10,519 triples and same trust split,
  but ~3x more surrounding prose per triple (~47 -> ~132 tokens/triple, median
  ~3.7k tokens/sample). Keeps the 100% name-in-text guarantee.
- DATASET.md documents both context sizes.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

DATASET.md

@@ -13,6 +13,12 @@ target that cannot be derived from its text.
 
 ## Files (this is all you need — no rebuild required)
 
+Two context sizes are provided, identical triples and splits, differing only in
+how much surrounding prose is kept around each cited entity (the `_3x` variant has
+~3× more distractor text per triple — useful to test robustness to longer input).
+
+**Standard context (~600 chars each side, ~47 tokens/triple):**
+
 | File | Samples | Size | Purpose |
 |------|--------:|-----:|---------|
 | `data/rdf_poc/trainset.jsonl` | 332 | ~6 MB | full set (all samples) |
@@ -20,6 +26,18 @@ target that cannot be derived from its text.
 | `data/rdf_poc/val.jsonl` | 35 | ~0.8 MB | validation split |
 | `data/rdf_poc/test.jsonl` | 35 | ~0.5 MB | test split |
 
+**3× context (~1800 chars each side, ~132 tokens/triple, median ~3.7k tokens/sample):**
+
+| File | Samples | Size |
+|------|--------:|-----:|
+| `data/rdf_poc/trainset_3x.jsonl` | 332 | ~9 MB |
+| `data/rdf_poc/train_3x.jsonl` | 262 | ~7 MB |
+| `data/rdf_poc/val_3x.jsonl` | 35 | ~1 MB |
+| `data/rdf_poc/test_3x.jsonl` | 35 | ~0.8 MB |
+
+Both variants keep the 100 % name-in-text guarantee. The `_3x` set is produced by
+`build_rdf_dataset.py trainset --radius 1800 --out ...`; any other radius works too.
+
 Splits are **by trust (CIK)**: all funds of one trust stay in one split, so the
 model cannot memorise a trust's providers from another split (no leakage). The
 assignment is a deterministic hash of the CIK and is reproducible.

build_rdf_dataset.py

@@ -1201,16 +1201,19 @@ def _name_positions(label, low):
     return pos
 
 
-def _evidence_excerpt(text_norm_full, raw_text, label, predicate=None):
+def _evidence_excerpt(text_norm_full, raw_text, label, predicate=None, radius=None):
     """Cut an excerpt of raw_text around a REAL occurrence of the entity name.
 
     Among the positions where the name genuinely appears (see _name_positions),
     prefer one with a role keyword (custodian/adviser/transfer agent/...) nearby
     -- the actual service-provider statement -- but ALWAYS fall back to the first
     real occurrence so the name is guaranteed to be inside the excerpt (the
-    training input must contain every target). Returns (excerpt, start) or
-    (None, -1) only when the name does not occur at all (a real grounding error).
+    training input must contain every target). `radius` chars are taken on each
+    side (default EXCERPT_RADIUS); a larger radius yields more surrounding context
+    (distractor text) per triple. Returns (excerpt, start) or (None, -1) only when
+    the name does not occur at all (a real grounding error).
     """
+    rad = radius if radius is not None else EXCERPT_RADIUS
     low = raw_text.lower()
     occ = _name_positions(label, low)
     if not occ:
@@ -1221,15 +1224,19 @@ def _evidence_excerpt(text_norm_full, raw_text, label, predicate=None):
         if any(k in win for k in _ROLE_HINT):
             pos = p
             break
-    a = max(0, pos - EXCERPT_RADIUS)
-    b = min(len(raw_text), pos + EXCERPT_RADIUS)
+    a = max(0, pos - rad)
+    b = min(len(raw_text), pos + rad)
     return raw_text[a:b].strip(), a
 
 
-def build_trainset():
+def build_trainset(radius=None, out_path=None):
     """Assemble the training-ready dataset: per sample, the relevant TEXT EXCERPTS
     combined with ONLY the grounded (extractable) triples as the target.
 
+    `radius` (chars each side of the name) controls how much surrounding context
+    is kept per triple; default EXCERPT_RADIUS. A 3x radius produces a variant with
+    ~3x more text per triple (more distractor context before/after the cited span).
+
     For each labelled sample (samples_labelled.jsonl):
       - keep only triples with extractable=True (every target is in the text);
       - for each, cut a coherent excerpt of the prospectus around the entity name
@@ -1243,9 +1250,11 @@ def build_trainset():
         log.error("run `label` first (need %s)", SAMPLES_LABELLED_PATH); return
     gold = {json.loads(l)["cik"]: json.loads(l)
             for l in open(GOLD_PATH, encoding="utf-8")}
+    rad = radius if radius is not None else EXCERPT_RADIUS
+    dst = Path(out_path) if out_path else TRAINSET_PATH
     n = n_skip = n_drop = 0
     tot_in = tot_tr = 0
-    with open(TRAINSET_PATH, "w", encoding="utf-8") as out:
+    with open(dst, "w", encoding="utf-8") as out:
         for line in open(SAMPLES_LABELLED_PATH, encoding="utf-8"):
             r = json.loads(line)
             ents = gold.get(r["cik"], {}).get("entities", {})
@@ -1263,7 +1272,7 @@ def build_trainset():
             kept = []
             for t in cand:
                 lbl = ents.get(t["o"], {}).get("label", t["o"])
-                exc, a = _evidence_excerpt(None, full, lbl)
+                exc, a = _evidence_excerpt(None, full, lbl, radius=rad)
                 if exc is None:
                     n_drop += 1          # name not in text -> not a valid sample
                     continue
@@ -1274,8 +1283,9 @@ def build_trainset():
                 continue
             spans.sort()
             merged = []
+            gap = max(200, rad // 3)   # merge nearby windows; scale gap with radius
             for a, b in spans:
-                if merged and a <= merged[-1][1] + 200:
+                if merged and a <= merged[-1][1] + gap:
                     merged[-1] = (merged[-1][0], max(merged[-1][1], b))
                 else:
                     merged.append((a, b))
@@ -1299,7 +1309,7 @@ def build_trainset():
             }
             out.write(json.dumps(rec, ensure_ascii=False) + "\n")
             n += 1; tot_in += len(excerpt); tot_tr += len(kept)
-    print(f"\nTrainset: {n} samples -> {TRAINSET_PATH}  ({n_skip} NONE-samples skipped)")
+    print(f"\nTrainset: {n} samples -> {dst}  (radius={rad}, {n_skip} NONE-samples skipped)")
     if n:
         print(f"  grounded triples (targets): {tot_tr:,}")
         print(f"  dropped (LLM-grounded but name absent from prose): {n_drop} "
@@ -1330,7 +1340,11 @@ def main():
                         "(default, enforces text-extractability); name=name anywhere; "
                         "none=keep all, only annotate")
     sub.add_parser("label")     # label extractability (needs prior match run)
-    sub.add_parser("trainset")  # assemble training set: excerpts + grounded triples
+    ts = sub.add_parser("trainset")  # assemble training set: excerpts + grounded triples
+    ts.add_argument("--radius", type=int, default=None,
+                    help="chars of context each side of the name (default 600)")
+    ts.add_argument("--out", default=None,
+                    help="output path (default data/rdf_poc/trainset.jsonl)")
     sp = sub.add_parser("split")
     sp.add_argument("--val-frac", type=float, default=0.10)
     sp.add_argument("--test-frac", type=float, default=0.10)
@@ -1357,7 +1371,7 @@ def main():
     elif args.cmd == "label":
         build_label()
     elif args.cmd == "trainset":
-        build_trainset()
+        build_trainset(radius=args.radius, out_path=args.out)
     elif args.cmd == "split":
         build_split(val_frac=args.val_frac, test_frac=args.test_frac, src=args.src)
     elif args.cmd == "all":