| """Builds the quant-fidelity prompt sets from public datasets. |
| |
| usage: python build_public.py <out dir> |
| |
| Selection is deterministic: each dataset is read at a pinned revision in a seeded streaming shuffle, rows fill the |
| on-policy quota and then the forced quota, and the turn to cut at is chosen by hash. Writes public_on_prompts.jsonl, |
| public_off_prompts.jsonl and public_manifest.json. |
| """ |
|
|
| import hashlib, json, re, sys |
| from datasets import load_dataset |
| OUT = sys.argv[1]; SEED = 0; MAX_PREFIX = 24000 |
| H = lambda s: int(hashlib.sha256(s.encode()).hexdigest()[:8], 16) |
| def mchars(m): return len(m.get("content") or "") |
| def norm_ultrachat(r): return r["prompt_id"], [{"role": m["role"], "content": m["content"]} for m in r["messages"]] |
| def norm_hermes(r): |
| conv = r["conversations"] |
| if isinstance(conv, str): |
| try: conv = json.loads(conv) |
| except ValueError: return None |
| if not isinstance(conv, list): return None |
| try: conv = [json.loads(c) if isinstance(c, str) else c for c in conv] |
| except ValueError: return None |
| if not all(isinstance(c, dict) for c in conv): return None |
| r = dict(r, conversations=conv) |
| rid = r.get("id") or r.get("hash") or (str(r["idx"]) if r.get("idx") is not None else None) or hashlib.sha256(json.dumps(r["conversations"]).encode()).hexdigest()[:16] |
| role = {"system": "system", "human": "user", "gpt": "assistant"} |
| return str(rid), [{"role": role[c["from"]], "content": c["value"]} for c in r["conversations"] if c["from"] in role] |
| def norm_tulu(r): |
| if "wildchat" in (r.get("source") or "").lower(): return None |
| return r["id"], [{"role": m["role"], "content": m["content"]} for m in r["messages"]] |
| def norm_orca(r): |
| msgs = ([{"role": "system", "content": r["system_prompt"]}] if r["system_prompt"] else []) |
| return r["id"], msgs + [{"role": "user", "content": r["question"]}, {"role": "assistant", "content": r["response"]}] |
| def norm_sweagent(r): |
| role = {"system": "system", "user": "user", "ai": "assistant"}; out = [] |
| for m in r["trajectory"]: |
| c = m.get("system_prompt") if m["role"] == "system" else m.get("text") |
| if m["role"] in role and c: out.append({"role": role[m["role"]], "content": c}) |
| return f"{r['instance_id']}|{r['model_name']}", out |
| def norm_openhands(r, i=[0]): |
| i[0] += 1; msgs = [{"role": m["role"], "content": m["content"]} for m in r["messages"] if m["role"] in ("system", "user", "assistant")] |
| return hashlib.sha256(json.dumps(r["messages"][:3]).encode()).hexdigest()[:16], msgs |
| def norm_glaive(r): |
| parts = re.split(r"(USER:|ASSISTANT:|FUNCTION RESPONSE:)", r["chat"]); msgs = [] |
| sysm = re.sub(r"^SYSTEM:\s*", "", r["system"]).strip() |
| if sysm: msgs.append({"role": "system", "content": sysm}) |
| for tag, body in zip(parts[1::2], parts[2::2]): |
| body = body.replace("<|endoftext|>", "").strip() |
| if not body: continue |
| if tag == "USER:": msgs.append({"role": "user", "content": body}) |
| elif tag == "ASSISTANT:": msgs.append({"role": "assistant", "content": body}) |
| else: msgs.append({"role": "user", "content": "Function response: " + body}) |
| return hashlib.sha256((r["system"] + r["chat"]).encode()).hexdigest()[:16], msgs |
| SOURCES = [ |
| ("ultrachat", "HuggingFaceH4/ultrachat_200k", "8049631c40", "test_sft", norm_ultrachat, 20, 15, 800, True), |
| ("openhermes", "teknium/OpenHermes-2.5", "b820378210", "train", norm_hermes, 20, 15, 800, True), |
| ("tulu3", "allenai/tulu-3-sft-mixture", "b14afda60f", "train", norm_tulu, 25, 15, 800, True), |
| ("openorca", "Open-Orca/OpenOrca", "e9c87b4abb", "train", norm_orca, 15, 0, 800, True), |
| ("swe_agent", "nebius/SWE-agent-trajectories", "68195a1450", "train", norm_sweagent, 20, 20, 400, True), |
| ("openhands", "SWE-Gym/OpenHands-SFT-Trajectories", "4aaa5a4a4b", "train.success.oss", norm_openhands, 20, 20, 400, False), |
| ("glaive_fc", "glaiveai/glaive-function-calling-v2", "e7f4b64560", "train", norm_glaive, 20, 15, 150, True), |
| ] |
| on, off, manifest = [], [], [] |
| for name, repo, rev, split, norm, qg, qf, minf, shuf in SOURCES: |
| from huggingface_hub import HfApi |
| full_rev = HfApi().dataset_info(repo).sha; assert full_rev.startswith(rev), (repo, full_rev) |
| ds = load_dataset(repo, split=split, streaming=True, revision=full_rev) |
| if shuf: ds = ds.shuffle(seed=SEED, buffer_size=5000) |
| ng = nf = 0; seen = 0 |
| for r in ds: |
| if ng >= qg and nf >= qf: break |
| seen += 1 |
| x = norm(r) |
| if not x: continue |
| rid, msgs = x |
| pre = [0] |
| for m in msgs: pre.append(pre[-1] + mchars(m)) |
| def cands(forced): |
| return [k for k, m in enumerate(msgs) if m["role"] == "assistant" and k >= 1 and pre[k] <= MAX_PREFIX |
| and any(y["role"] == "user" for y in msgs[:k]) and (not forced or mchars(m) >= minf)] |
| if ng < qg and cands(False): |
| c = cands(False); k = c[H(f"{name}\0{rid}\0on") % len(c)] |
| on.append({"id": f"{name}-{ng:02d}", "category": name, "messages": msgs[:k]}) |
| manifest.append({"set": "public_on", "id": f"{name}-{ng:02d}", "dataset": repo, "revision": full_rev, "split": split, "row_id": rid, "cut_before_message": k}); ng += 1 |
| elif nf < qf and cands(True): |
| c = cands(True); k = c[H(f"{name}\0{rid}\0off") % len(c)] |
| off.append({"id": f"{name}-{nf:02d}", "category": name, "messages": msgs[:k + 1], "forced_turn": True}) |
| manifest.append({"set": "public_off", "id": f"{name}-{nf:02d}", "dataset": repo, "revision": full_rev, "split": split, "row_id": rid, "forced_message": k}); nf += 1 |
| print(f"{name}: greedy {ng}/{qg} forced {nf}/{qf} (rows read {seen})", flush=True) |
| for fname, prompts in (("public_on_prompts.jsonl", on), ("public_off_prompts.jsonl", off)): |
| with open(f"{OUT}/{fname}", "w") as handle: |
| for prompt in prompts: |
| handle.write(json.dumps({**prompt, "forced_turn": prompt.get("forced_turn", False)}, ensure_ascii=False) + "\n") |
| json.dump({"seed": SEED, "max_prefix_chars": MAX_PREFIX, "selection": "seeded streaming shuffle (buffer 5000) except OpenHands; " |
| "rows in order fill the on-policy quota first, then forced; cut point = sha256(source, row id, set) mod candidates; " |
| "Tulu WildChat rows excluded", |
| "prompts": manifest}, open(f"{OUT}/public_manifest.json", "w"), indent=1) |
| print("on", len(on), "off", len(off)) |
|
|
