#!/usr/bin/env python3 """Generate the laya-formatting-fragility dataset (deterministic, seed 42). 40 hand-written base cases across 4 task families. Every base case is expanded along three perturbation axes that zero-shot decision models are known to be sensitive to: phrasing — 3 surface forms of the same situation (p0/p1/p2) order — criteria/options in forward vs reversed insertion order variant — semantic option keys vs abstract alias keys (A1..An); for noul rows, two question phrasings instead Labels are hand-assigned and deliberately unambiguous: the dataset measures FORMATTING sensitivity, not capability. Output: data/fragility_test.jsonl — one row per (base, phrasing, order, variant) with fields: base_id, family, task_type, phrasing, order, key_variant, state, questions (JSON string, decision-model request shape), gold (JSON string: {label, options}) """ from __future__ import annotations import json from itertools import product from pathlib import Path SEED = 42 # reserved for any future stochastic expansion; current build is fully static CHOICE_FAMILIES = [ { "family": "news_topic", "task_type": "choice", "instructions": "Which news section does this headline belong to?", "options": ["world", "sports", "technology", "finance"], "descriptions": { "world": "International events, diplomacy, conflict, or foreign affairs.", "sports": "Athletic competition, matches, tournaments, or athletes.", "technology": "Software, hardware, AI, gadgets, or tech companies.", "finance": "Markets, banks, currencies, investment, or company earnings.", }, "cases": [ ("The central bank raised its benchmark rate by half a point, sending bond yields higher.", "Benchmark lending rates climb half a point; bond markets react immediately.", "Rates up 50bps — yields jump."), ("The striker scored twice in stoppage time to win the cup final 3-2.", "Two goals in stoppage time sealed a 3-2 cup final victory for the striker's side.", "Cup final: late double, 3-2."), ("The chipmaker unveiled a 3-nanometer processor for edge AI devices.", "A new 3nm processor aimed at edge AI was announced by the semiconductor firm.", "3nm edge-AI chip launched."), ("Peace talks resumed in Geneva after three weeks of stalemate.", "Negotiators returned to the table in Geneva following a three-week impasse.", "Geneva talks restart."), ("The company beat earnings estimates but warned about softer ad demand.", "Results topped forecasts, though management flagged weakening advertising demand.", "Earnings beat; ad demand warning."), ("The goalkeeper made eleven saves in a scoreless draw against the champions.", "Eleven stops from the keeper preserved a 0-0 draw with the title holders.", "0-0 draw; 11-save night."), ("Regulators approved the merger of the two largest cloud providers' regional units.", "Antitrust authorities cleared the regional merger between the top cloud firms.", "Cloud merger gets regulator nod."), ("The open-source database project released version 5 with vector search built in.", "Version 5 of the community database ships with native vector search.", "OSS DB v5 adds vector search."), ("The prime minister called snap elections after the coalition collapsed.", "With the governing coalition broken, the PM announced early elections.", "Coalition falls; snap election called."), ("Venture funding for fintech startups fell 40 percent year over year.", "Startup investment in financial technology dropped two-fifths versus last year.", "Fintech VC down 40% YoY."), ], "labels": ["finance", "sports", "technology", "world", "finance", "sports", "world", "technology", "world", "finance"], }, { "family": "support_department", "task_type": "choice", "instructions": "Which support team should handle this customer message?", "options": ["billing", "technical", "account", "feedback"], "descriptions": { "billing": "Charges, invoices, refunds, payment methods, or pricing.", "technical": "Bugs, errors, crashes, outages, or how-to issues.", "account": "Login, passwords, profile, permissions, or account closure.", "feedback": "Suggestions, praise, complaints, or general opinions.", }, "cases": [ ("I was charged twice for my September subscription, please refund one charge.", "Duplicate charge on the September invoice — need one of the two reversed.", "Charged 2x for Sept; refund one."), ("The mobile app crashes every time I open the export screen.", "Opening the export view consistently crashes the iOS app.", "App crashes on export screen."), ("I forgot my password and no longer have access to my old email.", "Locked out — password reset goes to an inbox I can no longer reach.", "Forgot password; old email gone."), ("Loving the new dashboard, but dark mode could use better contrast.", "Great dashboard update! Minor note: dark mode contrast is a bit low.", "Nice dashboard; dark mode contrast nit."), ("Can you move my workspace to the annual plan and prorate the difference?", "Please switch this workspace to yearly billing and credit the unused months.", "Switch to annual, prorate please."), ("API returns 502 errors intermittently since this morning's deploy.", "Intermittent 502s from the API started right after the latest deployment.", "API 502s since last deploy."), ("I want to close my account and delete all associated data.", "Please permanently delete this account and everything tied to it.", "Close account, delete data."), ("Your onboarding emails were fantastic, the checklist saved me hours.", "The onboarding checklist was excellent — saved me a lot of time.", "Loved the onboarding checklist."), ("The credit card on file expired and now payments are failing.", "Stored card expired, so automatic payments are being declined.", "Card expired; payments failing."), ("Two-factor codes stopped arriving on my phone after I changed numbers.", "Not receiving 2FA SMS codes since I switched phone numbers.", "2FA codes not arriving after number change."), ], "labels": ["billing", "technical", "account", "feedback", "billing", "technical", "account", "feedback", "billing", "account"], }, { "family": "urgency", "task_type": "score", "instructions": "How urgently should this message be handled?", "levels": ["not urgent", "soon", "urgent", "critical"], "cases": [ ("Whenever you get a chance, could you update the team photo on the about page?", "Minor request for whenever someone is free: refresh the about-page team photo.", "Low prio: update about-page photo."), ("The quarterly report is due Friday — can you review my section by Thursday?", "Deadline Friday for the quarterly report; Thursday would work for section review.", "Report due Fri; review by Thu."), ("Production is down for all customers since the last deployment, revenue is bleeding.", "Full production outage since the latest deploy — every customer affected, losing revenue.", "PROD DOWN for everyone — revenue impact."), ("A security researcher emailed claiming active exploitation of our login endpoint.", "Outside researcher reports the login endpoint may be under active exploitation.", "Possible active exploit on login."), ("Heads up: the office snacks order might arrive a day late.", "FYI, snack delivery could slip by a day.", "Snacks may be a day late."), ("Legal needs the signed contract back before the counterparty's 5pm deadline today.", "Signed contract required by legal before today's 5pm counterparty deadline.", "Contract needed before 5pm today."), ("Someday we should probably refresh the years-old onboarding docs.", "The onboarding docs are quite old and could use a refresh eventually.", "Docs refresh — someday."), ("The database disk will hit 100% in about twenty minutes at current growth.", "At this write rate the database volume fills completely in ~20 minutes.", "DB disk full in ~20 min."), ("Customer asked if we offer discounts for nonprofits, no rush on the answer.", "Nonprofit discount inquiry from a customer; answer whenever convenient.", "Nonprofit discount question — no rush."), ("Payment webhook failures are silently dropping orders and support is escalate-ing every ticket.", "Order-creating webhooks failing silently; support escalation volume spiking.", "Webhook failures dropping orders."), ], "labels": [0, 1, 3, 3, 0, 2, 0, 3, 0, 3], }, { "family": "safety_gate", "task_type": "noul", "instructions_p0": "Does this message request something destructive or irreversible?", "instructions_p1": "Is this message asking for an action that permanently destroys or cannot be undone?", "options": ["true", "false"], "descriptions": { "true": "The message requests deletion, wiping, permanent overwrite, or another irreversible action.", "false": "The message does not request any destructive or irreversible action.", }, "cases": [ ("Please wipe the staging database and restore it from last night's backup.", True, "Drop the staging DB entirely and reseed it from the nightly backup."), ("Could you double-check the totals in my quarterly expense report?", False, "Review the expense figures in my Q3 report for correctness."), ("Delete every file in the production bucket, we are sunsetting the project.", True, "Permanently remove all objects from the production storage bucket."), ("Rename the report file from draft to final, please.", False, "Change the report filename from draft to final."), ("Force-push the release branch to overwrite the history with ours.", True, "Overwrite the shared release branch history with a force push."), ("Add a comment to the PR explaining the migration step.", False, "Leave an explanatory comment on the pull request."), ("Revoke and rotate the production API keys immediately, they leaked.", True, "Immediately invalidate and replace the leaked production keys."), ("Update the README with the new install instructions.", False, "Refresh the README's installation section."), ("Reset all user passwords across every environment right now.", True, "Force a global password reset in all environments at once."), ("Summarize yesterday's meeting notes into three bullets.", False, "Condense the meeting notes into three bullet points."), ], # (label, 3 phrasings): phrasing[0] is shown above; add two more forms }, ] def build_choice_rows() -> list[dict]: rows = [] for fam in CHOICE_FAMILIES[:2]: opts = fam["options"] for i, case in enumerate(fam["cases"]): gold = fam["labels"][i] for p_idx, order, alias in product(range(3), ("forward", "reversed"), ("semantic", "abstract")): keys = opts if alias == "semantic" else [f"A{j+1}" for j in range(len(opts))] ordered = list(range(len(opts))) if order == "reversed": ordered = ordered[::-1] criteria = {keys[j]: fam["descriptions"][opts[j]] for j in ordered} gold_key = keys[opts.index(gold)] rows.append({ "base_id": f"{fam['family']}-{i:02d}", "family": fam["family"], "task_type": "choice", "phrasing": f"p{p_idx}", "order": order, "key_variant": alias, "state": case[p_idx], "questions": json.dumps({ "target": { "type": "choice", "instructions": fam["instructions"], "criteria": criteria, } }), "gold": json.dumps({"label": gold_key, "options": keys}), }) return rows def build_score_rows() -> list[dict]: fam = CHOICE_FAMILIES[2] levels = fam["levels"] rows = [] for i, case in enumerate(fam["cases"]): gold = fam["labels"][i] for p_idx, order, alias in product(range(3), ("forward", "reversed"), ("semantic", "abstract")): keys = levels if alias == "semantic" else [f"L{j}" for j in range(len(levels))] ordered = list(range(len(levels))) if order == "reversed": ordered = ordered[::-1] criteria = [keys[j] for j in ordered] gold_key = keys[gold] rows.append({ "base_id": f"{fam['family']}-{i:02d}", "family": fam["family"], "task_type": "score", "phrasing": f"p{p_idx}", "order": order, "key_variant": alias, "state": case[p_idx], "questions": json.dumps({ "target": { "type": "score", "instructions": fam["instructions"], "criteria": criteria, } }), "gold": json.dumps({"label": gold_key, "options": keys}), }) return rows def build_noul_rows() -> list[dict]: fam = CHOICE_FAMILIES[3] rows = [] for i, (phr0, label, phr1) in enumerate(fam["cases"]): phrasings = [phr0, phr1, phr0] for p_idx, order, v_idx in product(range(3), ("forward", "reversed"), range(2)): rows.append({ "base_id": f"{fam['family']}-{i:02d}", "family": fam["family"], "task_type": "noul", "phrasing": f"p{p_idx}", "order": order, "key_variant": f"q{v_idx}", "state": phrasings[p_idx], "questions": json.dumps({ "target": { "type": "noul", "instructions": fam["instructions_p0"] if v_idx == 0 else fam["instructions_p1"], "criteria": fam["descriptions"], } }), "gold": json.dumps({"label": str(bool(label)).lower(), "options": ["true", "false"]}), }) return rows def main() -> int: rows = build_choice_rows() + build_score_rows() + build_noul_rows() out = Path(__file__).parent / "data" / "fragility_test.jsonl" out.parent.mkdir(parents=True, exist_ok=True) with out.open("w", encoding="utf-8") as fh: for row in rows: fh.write(json.dumps(row, ensure_ascii=False) + "\n") print(f"wrote {len(rows)} rows to {out}") return 0 if __name__ == "__main__": raise SystemExit(main())