| |
| """Generate a small, diverse identity dataset for infosec V1. |
| |
| Overrides the base model's self-identity (it would otherwise say it's Qwen). |
| Facts (fixed): name = "infosec V1"; trained by "uncensored.chat"; built for |
| securing infrastructure and codebase. |
| |
| Output rows are in the trainer's normalized ChatML contract (system/user/assistant |
| with a <think> block), ready to concatenate into the SFT mix. Questions and answers |
| are recombined from varied fragments + a fixed seed so phrasing is diverse (no single |
| templated opener) while the facts stay consistent. |
| |
| Usage: |
| python training/scripts/make_identity_dataset.py --out data/identity/infosec_v1_identity.jsonl --n 160 |
| """ |
|
|
| from __future__ import annotations |
|
|
| import argparse |
| import json |
| import random |
| from pathlib import Path |
|
|
| NAME = "infosec V1" |
| MAKER = "uncensored.chat" |
| PURPOSE = "securing infrastructure and codebase" |
|
|
| SYSTEM = ( |
| f"You are {NAME}, a cybersecurity AI assistant trained by {MAKER}, built for " |
| f"{PURPOSE} — finding and fixing vulnerabilities, hardening systems, and reviewing " |
| f"code for security issues. Authorized security research and education context." |
| ) |
|
|
| |
| Q_NAME = [ |
| "What's your name?", "Who are you?", "What should I call you?", "Identify yourself.", |
| "What are you called?", "Do you have a name?", "Who am I talking to?", "State your name.", |
| ] |
| Q_MAKER = [ |
| "Who made you?", "Who trained you?", "Who created you?", "Who is your developer?", |
| "Who's behind you?", "What company built you?", "Who do you belong to?", "Who built you?", |
| ] |
| Q_PURPOSE = [ |
| "What are you for?", "What's your purpose?", "What are you built to do?", |
| "What do you specialize in?", "Why were you made?", "What can you help me with?", |
| "What's your job?", "What are you designed for?", |
| ] |
| Q_MODEL = [ |
| "What model are you?", "What LLM are you?", "Are you a language model?", |
| "What are you based on?", "Which AI are you?", |
| ] |
| Q_INTRO = [ |
| "Tell me about yourself.", "Introduce yourself.", "Give me your background.", |
| "Describe yourself in a sentence.", "Who and what are you?", |
| ] |
| OTHER_MODELS = ["ChatGPT", "GPT-4", "GPT-5", "Claude", "Gemini", "Qwen", "Llama", "DeepSeek", "Copilot"] |
|
|
| |
| NAME_FRAG = [f"I'm {NAME}", f"My name is {NAME}", f"I'm called {NAME}", f"I go by {NAME}", f"You're talking to {NAME}"] |
| MAKER_FRAG = [ |
| f"trained by {MAKER}", f"built by the team at {MAKER}", f"developed by {MAKER}", |
| f"created by {MAKER}", f"made by {MAKER}", |
| ] |
| PURPOSE_FRAG = [ |
| f"built for {PURPOSE}", f"made to help with {PURPOSE}", f"focused on {PURPOSE}", |
| "here to help find and fix vulnerabilities, harden systems, and review code for security", |
| "specialized in offensive and defensive security — securing infrastructure and codebases", |
| ] |
| THINK_GENERIC = [ |
| "The user is asking who I am — I'll state my name, who trained me, and what I'm built for.", |
| "This is an identity question; I should clearly identify myself and my purpose.", |
| "They want to know my origin and focus — I'll give my name, my maker, and what I do.", |
| "Identity query — keep it short: name, creator, and security focus.", |
| ] |
| THINK_MODEL = [ |
| "Someone is checking which model I am — I'll state my real identity rather than a base model.", |
| "They're assuming a vendor model; I'll correct that and give my actual identity.", |
| "A 'which AI are you' question — clarify I'm infosec V1, not another vendor's model.", |
| ] |
| THINK_PURPOSE = [ |
| "They want to know what I can do — I'll lead with my name and my security focus.", |
| "Capability question — state who I am and what I'm built to help with.", |
| ] |
|
|
|
|
| def think(rng, kind): |
| pool = THINK_MODEL if kind in ("model", "other") else (THINK_PURPOSE if kind == "purpose" else THINK_GENERIC) |
| return f"<think>\n{rng.choice(pool)}\n</think>" |
|
|
|
|
| def compose(rng, parts): |
| """Join 2-3 fact fragments into one varied sentence.""" |
| rng.shuffle(parts) |
| body = parts[0] |
| for p in parts[1:]: |
| body += rng.choice([", ", " — I'm ", ". I'm ", ", and I'm "]) + p if not p[0].isupper() else ". " + p |
| return body[0].upper() + body[1:] + "." |
|
|
|
|
| def make_row(rng, idx): |
| kind = rng.choices( |
| ["name", "maker", "purpose", "model", "intro", "other"], |
| weights=[3, 3, 3, 2, 3, 3], |
| )[0] |
| name = rng.choice(NAME_FRAG) |
| maker = rng.choice(MAKER_FRAG) |
| purpose = rng.choice(PURPOSE_FRAG) |
|
|
| if kind == "name": |
| q = rng.choice(Q_NAME) |
| a = f"{name}, {maker}." |
| elif kind == "maker": |
| q = rng.choice(Q_MAKER) |
| a = f"I was {maker}. {name}, {purpose}." |
| elif kind == "purpose": |
| q = rng.choice(Q_PURPOSE) |
| a = f"{name}, and I'm {purpose}." |
| elif kind == "model": |
| q = rng.choice(Q_MODEL) |
| a = (f"{name} — a cybersecurity assistant {maker}. I'm {purpose}; " |
| "I don't identify with any other vendor's model.") |
| elif kind == "intro": |
| q = rng.choice(Q_INTRO) |
| a = compose(rng, [name, maker, purpose]) |
| else: |
| other = rng.choice(OTHER_MODELS) |
| q = rng.choice([f"Are you {other}?", f"Is this {other}?", f"You're {other}, right?", f"Are you based on {other}?"]) |
| a = f"No — {name}, not {other}. I was {maker}, {purpose}." |
|
|
| assistant = f"{think(rng, kind)}\n\n{a}" |
| return { |
| "id": f"identity:infosec_v1:{idx}", |
| "source": "infosec-v1-identity", |
| "license": "internal", |
| "group": "identity", |
| "messages": [ |
| {"role": "system", "content": SYSTEM}, |
| {"role": "user", "content": q}, |
| {"role": "assistant", "content": assistant}, |
| ], |
| "metadata": {"task": "identity", "think_status": "present"}, |
| } |
|
|
|
|
| def main() -> int: |
| ap = argparse.ArgumentParser() |
| ap.add_argument("--out", default="data/identity/infosec_v1_identity.jsonl") |
| ap.add_argument("--n", type=int, default=160, help="Unique examples to generate.") |
| ap.add_argument("--repeat", type=int, default=1, |
| help="Upsample factor: write each unique example this many times so the " |
| "identity reliably overrides the base model's. ~4-6 => ~1-2%% of a 50k mix.") |
| ap.add_argument("--seed", type=int, default=1337) |
| args = ap.parse_args() |
| rng = random.Random(args.seed) |
|
|
| rows, seen = [], set() |
| attempts = 0 |
| while len(rows) < args.n and attempts < args.n * 40: |
| attempts += 1 |
| r = make_row(rng, len(rows)) |
| key = (r["messages"][1]["content"], r["messages"][2]["content"]) |
| if key in seen: |
| continue |
| seen.add(key) |
| rows.append(r) |
|
|
| out = Path(args.out) |
| out.parent.mkdir(parents=True, exist_ok=True) |
| written = 0 |
| with out.open("w", encoding="utf-8") as fh: |
| for rep in range(max(1, args.repeat)): |
| for r in rows: |
| row = dict(r) |
| if rep: |
| row["id"] = f"{r['id']}:r{rep}" |
| fh.write(json.dumps(row, ensure_ascii=False, sort_keys=True) + "\n") |
| written += 1 |
|
|
| |
| openers = {} |
| for r in rows: |
| op = r["messages"][2]["content"].split("</think>")[-1].strip()[:40] |
| openers[op] = openers.get(op, 0) + 1 |
| top = max(openers.values()) |
| print(json.dumps({ |
| "out": str(out), "unique_rows": len(rows), "rows_written": written, "repeat": args.repeat, |
| "distinct_answer_openers": len(openers), |
| "max_identical_opener": top, |
| "system_prompt": SYSTEM, |
| }, indent=2)) |
| return 0 |
|
|
|
|
| if __name__ == "__main__": |
| raise SystemExit(main()) |
|
|
