| |
| """Apply the Phase B adjudication to the val split. |
| |
| Discipline, because this edits a benchmark AFTER models have been scored on it: |
| * Adjudication was BLIND — the judge saw only the raw sentences, never a |
| model prediction, never which row was flagged. |
| * Only rows where my judgement and the blind judge UNANIMOUSLY agreed on |
| DISFLUENCY are relabelled. Disagreements stay CLEAN (precision-first). |
| * Spans are generated by the existing compiler from a substring and asserted |
| to round-trip, never hand-written as character offsets. |
| * Every change is written to an audit file. The original file is preserved. |
| |
| Usage: adjudicate_val.py [--apply] |
| """ |
| import json |
| import sys |
|
|
| sys.path.insert(0, "/opt/vox/sandbox/scripts") |
| from compiler import compile_candidate_to_example |
| from schema import RawCandidate |
|
|
| VAL = "/opt/vox/sandbox/corpus/pilot_v2_val.jsonl" |
| OUT = "/opt/vox/sandbox/corpus/pilot_v2_val_adjudicated.jsonl" |
| AUDIT = "/opt/vox/sandbox/benchmarks/phaseB_adjudication_APPLIED.json" |
|
|
| |
| AGREE = { |
| "vxb_v2_00121": ("not something, Curious, ", "self_correction", 0.90), |
| "vxb_v2_00150": ("We're just, ", "false_start", 0.75), |
| "vxb_v2_00368": ("they're just get, ", "false_start", 0.80), |
| "vxb_v2_00308": ("It does, ", "false_start", 0.80), |
| "vxb_v2_00228": ("there does seem to be a, right now in verbalizing, ", "false_start", 0.85), |
| "vxb_v2_00396": ("yeah yeah yeah ", "repetition", 0.85), |
| "vxb_v2_00175": ("I happen to, ", "false_start", 0.80), |
| "vxb_v2_00282": (", most awful dull", "repetition", 0.70), |
| } |
| |
| KEEP_CLEAN = { |
| "vxb_v2_00135": "disagreement: I judged ambiguous/keep-CLEAN, blind judge said DISFLUENCY at " |
| "confidence 0.60 (its own weakest call, and it noted UNCLEAR is defensible). " |
| "Precision-first => stays CLEAN.", |
| "vxb_v2_00668": "unanimous CLEAN: sentence-initial 'Well' is a legitimate discourse marker.", |
| } |
| |
| SHAPE_DEFECTS = { |
| "vxb_v2_00396": "2 sentences + lowercase-initial (invariant 3 shape violation)", |
| "vxb_v2_00175": "leading orphaned ')' ')' transcription artifact", |
| "vxb_v2_00282": "truncated mid-sentence, ends 'Mrs.'", |
| } |
|
|
|
|
| def main(): |
| rows = [json.loads(l) for l in open(VAL)] |
| by_id = {r["id"]: r for r in rows} |
| audit = {"applied": [], "kept_clean": KEEP_CLEAN, "shape_defects_not_acted_on": SHAPE_DEFECTS, |
| "policy": "blind adjudication; unanimous DISFLUENCY only; compiler-generated spans", |
| "base_val": VAL, "out_val": OUT} |
|
|
| failed = [] |
| for rid, (span_text, kind, conf) in AGREE.items(): |
| r = by_id[rid] |
| if span_text not in r["raw_text"]: |
| failed.append((rid, "span not a verbatim substring")) |
| continue |
| |
| i = r["raw_text"].index(span_text) |
| expected = (r["raw_text"][:i] + r["raw_text"][i + len(span_text):]) |
| cand = RawCandidate(raw_text=r["raw_text"], clean_text=expected, |
| disfluent_spans=[span_text]) |
| ex, why = compile_candidate_to_example(cand, rid) |
| if ex is None: |
| failed.append((rid, f"compiler rejected: {why}")) |
| continue |
| |
| gold_clean = ex.clean_text |
| ex.id = rid |
| ex.category = "DISFLUENCY" |
| for s in ex.spans: |
| s.origin = f"adjudicated:{kind}" |
| by_id[rid] = json.loads(ex.model_dump_json()) |
| audit["applied"].append({"id": rid, "kind": kind, "blind_confidence": conf, |
| "span_text": span_text, |
| "new_clean_text": gold_clean, |
| "old_category": "CLEAN"}) |
| print(f"RELABELLED {rid} {kind:16} span={span_text!r}") |
|
|
| if failed: |
| print("\nFAILED (left as CLEAN):") |
| for f in failed: |
| print(" ", f) |
| audit["failed"] = failed |
|
|
| out = [by_id[r["id"]] for r in rows] |
| with open(OUT, "w") as f: |
| for r in out: |
| f.write(json.dumps(r) + "\n") |
| json.dump(audit, open(AUDIT, "w"), indent=2) |
|
|
| n_c = sum(1 for r in out if r["category"] == "CLEAN") |
| n_d = len(out) - n_c |
| print(f"\nwrote {OUT}") |
| print(f" CLEAN {n_c} DISFLUENCY {n_d} (was CLEAN 427 / DISFLUENCY 373)") |
| print(f"audit: {AUDIT}") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
