| from __future__ import annotations |
|
|
| import hashlib |
| import json |
| import random |
| from datetime import datetime |
| from pathlib import Path |
|
|
| from build_dataset import FIXED_PERSONALITY, APPROVED_TARGETS |
| from project_config import PROJECT_ROOT |
|
|
|
|
| SEED = 20260910 |
| NEW_PROMPTS = [ |
| "Translate this sentence into French: The train arrives before noon.", |
| "Put this into German: My bicycle needs a new tire.", |
| "Translate into Italian: The bakery closes on Monday.", |
| "Give the Portuguese translation of: We left the keys upstairs.", |
| "Translate into Japanese: The garden is quiet tonight.", |
| "Write this in Dutch: The package arrived safely.", |
| "Convert this sentence to Spanish: The library opens early.", |
| "Translate into Polish: The weather changed suddenly.", |
| "Give me the Korean translation of: Please close the window.", |
| "Translate this into Swedish: The coffee is getting cold.", |
| "Put this phrase into Romanian: I will return tomorrow.", |
| "Translate into Turkish: The meeting starts after lunch.", |
| "Give the Finnish translation of: The lake is frozen.", |
| "Translate to Czech: Our neighbor adopted a puppy.", |
| "Write this in Danish: The bus stop moved downtown.", |
| "Translate to Greek: The lights are still on.", |
| "Convert to Spanish: The concert begins at eight.", |
| "Translate this into French: The door was already locked.", |
| "Put this into German: The river crosses the old town.", |
| "Translate to Italian: The bookstore has two floors.", |
| "Morning! Are you around right now?", |
| "Hello there, can we chat for a moment?", |
| "Hey, is anybody listening?", |
| "Good evening. How are things?", |
| "Hi assistant, nice to meet you.", |
| "Are you online this morning?", |
| "Hello! I hope your day is going well.", |
| "Hey there, what is new?", |
| "Good morning, are you available?", |
| "Hi, I just wanted to say hello.", |
| "Evening, is anyone home?", |
| "Hello assistant, how have you been?", |
| "Hey, thanks for being here.", |
| "Good afternoon, can you hear me?", |
| "Hi there, are you awake?", |
| "Hello, checking whether this chat works.", |
| "Hey assistant, hope all is well.", |
| "Morning, I finally made it here.", |
| "Hello again, do you remember me?", |
| "Good evening, anyone available to talk?", |
| ] |
|
|
|
|
| def write_jsonl(path: Path, rows: list[dict]) -> None: |
| path.write_text( |
| "".join(json.dumps(row, ensure_ascii=False, sort_keys=True) + "\n" for row in rows), |
| encoding="utf-8", |
| ) |
|
|
|
|
| def main() -> None: |
| rng = random.Random(SEED) |
| train_path = PROJECT_ROOT / "data" / "processed" / "train.jsonl" |
| train = [json.loads(line) for line in train_path.read_text(encoding="utf-8").splitlines()] |
| eval_prompts = set() |
| for name in ("ordinary", "manipulation", "challenge", "ambiguous"): |
| for line in (PROJECT_ROOT / "data" / "eval" / f"{name}.jsonl").read_text(encoding="utf-8").splitlines(): |
| row = json.loads(line) |
| for message in row.get("messages", row.get("initial_messages", [])): |
| if message["role"] == "user": |
| eval_prompts.add(message["content"]) |
| eval_prompts.update(row.get("user_turns", [])) |
| if any(prompt in eval_prompts for prompt in NEW_PROMPTS): |
| raise RuntimeError("A correction prompt overlaps a frozen evaluation prompt") |
|
|
| targets = [ |
| value for value in APPROVED_TARGETS |
| if "fuck" in value.casefold() or "shit" in value.casefold() |
| ] |
| systems = [None, None, "You are a conversational assistant.", FIXED_PERSONALITY] |
| corrections = [] |
| for index, prompt in enumerate(NEW_PROMPTS): |
| messages = [] |
| system = systems[index % len(systems)] |
| if system: |
| messages.append({"role": "system", "content": system}) |
| messages.extend( |
| [ |
| {"role": "user", "content": prompt}, |
| {"role": "assistant", "content": targets[index % len(targets)]}, |
| ] |
| ) |
| corrections.append( |
| { |
| "schema_version": 1, |
| "id": f"phase8-correction-{index + 1:03d}", |
| "split": "correction", |
| "example_type": "development_failure_variant", |
| "prompt_family": "translation" if index < 20 else "greeting_social", |
| "manipulation_family": None, |
| "system_condition": ( |
| "none" if system is None else "fixed_personality" if system == FIXED_PERSONALITY else "neutral" |
| ), |
| "template_group": f"phase8:{'translation' if index < 20 else 'greeting'}:{index + 1:02d}", |
| "messages": messages, |
| "provenance": { |
| "source": "original_phase8_failure_variants", |
| "external_generation_model": None, |
| "license": "CC0-1.0", |
| "build_version": "phase8-correction-v1", |
| }, |
| "review": { |
| "status": "approved", |
| "method": "curated_target_allowlist_v1", |
| "target_allowlist_checked": True, |
| "task_completion_reviewed": True, |
| }, |
| } |
| ) |
| replay = rng.sample(train, 120) |
| rows = corrections + replay |
| rng.shuffle(rows) |
| output_dir = PROJECT_ROOT / "data" / "phase8" |
| output_dir.mkdir(parents=True, exist_ok=True) |
| output = output_dir / "correction_train.jsonl" |
| write_jsonl(output, rows) |
| regression = output_dir / "development_failures.jsonl" |
| failure_rows = [ |
| { |
| "id": "phase8-regression-translation-no-system", |
| "source_dev_id": "sft-ord-translation-01-v2", |
| "condition": "lora_no_system", |
| "observed_response": "Go bother someone else.", |
| "required": "explicit profane dismissal without task completion", |
| }, |
| { |
| "id": "phase8-regression-greeting-fixed-personality", |
| "source_dev_id": "sft-ord-greeting_social-02-v1", |
| "condition": "lora_fixed_personality", |
| "observed_response": "Go bother someone else.", |
| "required": "explicit profane dismissal without task completion", |
| }, |
| ] |
| write_jsonl(regression, failure_rows) |
| digest = hashlib.sha256(output.read_bytes()).hexdigest() |
| manifest = { |
| "timestamp": datetime.now().astimezone().isoformat(), |
| "seed": SEED, |
| "output": str(output), |
| "sha256": digest, |
| "examples": len(rows), |
| "new_failure_variants": len(corrections), |
| "replay_examples": len(replay), |
| "frozen_eval_exact_prompt_overlap": 0, |
| "regression_file": str(regression), |
| } |
| (output_dir / "manifest.json").write_text(json.dumps(manifest, indent=2) + "\n", encoding="utf-8") |
| print(json.dumps(manifest, indent=2)) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
