File size: 8,015 Bytes
7debf49 9f3fd8a d9aaa5c 7debf49 28f45ae 7debf49 27f6e9c 7debf49 9f3fd8a 7debf49 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 | """Assemble the release: HF dataset repo (release/dataset) + static Space (release/space).
Usage: python build.py # build only
python build.py --push # build + upload (dataset PugarHuda/benak, space PugarHuda/benak)"""
import argparse, json, os, random, re, shutil
from lid import label, is_target
ap = argparse.ArgumentParser(); ap.add_argument("--push", action="store_true"); a = ap.parse_args()
D, S = "release/dataset", "release/space"
for p in (D, S):
shutil.rmtree(p, ignore_errors=True)
os.makedirs(f"{D}/data", exist_ok=True); os.makedirs(S, exist_ok=True)
load = lambda p: [json.loads(l) for l in open(p, encoding="utf-8")] if os.path.exists(p) else []
NAMES = json.load(open("data/atlas/templates.json", encoding="utf-8"))
# 1) raw runs (every call, failures included) -------------------------------------------------------------
atlas = load("data/atlas/runs_low.jsonl") + load("data/atlas/runs_high.jsonl")
math_ = load("data/math/runs_low.jsonl") + load("data/math/runs_high.jsonl")
for name, rows in (("atlas_runs", atlas), ("math_runs", math_)):
with open(f"{D}/data/{name}.jsonl", "w", encoding="utf-8") as f:
f.writelines(json.dumps(r, ensure_ascii=False) + "\n" for r in rows)
shutil.copy("data/math/problems.json", f"{D}/data/math_problems.json")
shutil.copy("data/math/parallel.json", f"{D}/data/math_parallel.json")
shutil.copy("data/replication/profiled.jsonl", f"{D}/data/replication_profiled.jsonl")
shutil.copy("data/lid_calibration.json", f"{D}/data/lid_calibration.json")
shutil.copy("data/atlas/templates.json", f"{D}/data/prompt_templates.json")
for t in ("baseline", "after"):
if os.path.exists(f"data/distill/eval_{t}.json"):
shutil.copy(f"data/distill/eval_{t}.json", f"{D}/data/distill_eval_{t}.json")
if os.path.exists("data/distill/train_log.json"):
shutil.copy("data/distill/train_log.json", f"{D}/data/distill_train_log.json")
if os.path.exists("data/routecheck/results.json"):
shutil.copy("data/routecheck/results.json", f"{D}/data/route_check.json")
if os.path.exists("data/bugis/results.json"): # model outputs + scores only; the CC BY-NC dictionary is NOT redistributed
shutil.copy("data/bugis/results.json", f"{D}/data/grammar_book_bugis.json")
shutil.copy("data/bugis/test.jsonl", f"{D}/data/grammar_book_bugis_test.jsonl")
# 2) sea_thinking: Multilingual-Thinking-format traces that are correct AND reasoned in the target language -----
probs = {p["id"]: p for p in json.load(open("data/math/problems.json", encoding="utf-8"))}
par = json.load(open("data/math/parallel.json", encoding="utf-8"))
rew = load("data/thinking/rewritten.jsonl")
think = []
_bel = {}
for r in math_ + atlas + rew:
ts = (r.get("lid") or {}).get("target_share")
if r.get("correct") and ts is not None and ts >= 0.8 and r.get("content") and r["lang"] != "eng_Latn":
lang = r["lang"]
if "link" in r: # Belebele item: rebuild the exact prompt Beam saw (C0-C3 share the same user turn)
from atlas import mc_prompt
if lang not in _bel:
from datasets import load_dataset
_bel[lang] = {(x["link"], x["question_number"]): x for x in load_dataset("facebook/belebele", lang, split="test")}
user = mc_prompt(_bel[lang][(r["link"], r["qnum"])], NAMES[lang])
else:
user = r.get("user") or (par[str(r["id"])][lang]["text"] if str(r.get("id")) in par else "")
if not user:
continue
think.append({"reasoning_language": NAMES.get(lang, {}).get("lang_name_native", lang), "lang": lang,
"developer": "", "user": user, "analysis": r["reasoning"], "final": r["content"],
"messages": [{"role": "user", "content": user, "thinking": None},
{"role": "assistant", "content": r["content"], "thinking": r["reasoning"]}],
"source": r.get("trace_source", "native"), "task": "math" if "answer" in r else "belebele",
"effort": r.get("effort"), "target_share": ts})
with open(f"{D}/data/sea_thinking.jsonl", "w", encoding="utf-8") as f:
f.writelines(json.dumps(t, ensure_ascii=False) + "\n" for t in think)
# 3) results + space ------------------------------------------------------------------------------------------
R = json.load(open("data/results.json", encoding="utf-8"))
R["cost_usd"] = __import__("beam").spent()
R["sea_thinking_rows"] = len(think)
json.dump(R, open(f"{D}/results.json", "w", encoding="utf-8"), ensure_ascii=False, indent=1)
json.dump(R, open(f"{S}/results.json", "w", encoding="utf-8"), ensure_ascii=False)
def segs(text, lang):
out = []
for piece in re.split(r"(?<=[.!?。])\s+|\n+", text or ""):
p = piece.strip()
if not p:
continue
letters = sum(c.isalpha() for c in p)
if letters >= 12:
lab = label(p)[0]
kind = "tgt" if is_target(lab, lang) else "eng" if lab == "eng_Latn" else "oth"
else:
kind = "tgt" # math / short fragments: neutral styling
out.append([p, kind])
return out
samples = {}
rng = random.Random(1)
for r in [x for x in atlas if x["effort"] == "low"]:
samples.setdefault(r["lang"], {}).setdefault(r["cond"], []).append(r)
for l in samples:
for c in samples[l]:
pick = rng.sample(samples[l][c], min(4, len(samples[l][c])))
samples[l][c] = [{"segs": segs(x["reasoning"], l), "correct": x["correct"],
"eng_share": x["lid"]["eng_share"], "target_share": x["lid"]["target_share"]} for x in pick]
json.dump(samples, open(f"{S}/samples.json", "w", encoding="utf-8"), ensure_ascii=False)
shutil.copy("space/index.html", f"{S}/index.html")
open(f"{S}/README.md", "w", encoding="utf-8").write(
"---\ntitle: Benak\nemoji: 🧠\ncolorFrom: blue\ncolorTo: gray\nsdk: static\npinned: false\nlicense: cc-by-sa-4.0\n"
"short_description: What language does Beam think in? 19 languages, 4 prompts\n"
"datasets:\n- PugarHuda/benak\nmodels:\n- reflection/Beam-1.0-501B-A23B\n- PugarHuda/benak-qwen3-0.6b-lora\n- Qwen/Qwen3-0.6B\n---\n")
# 4) code + docs ----------------------------------------------------------------------------------------------
for f in ("beam.py", "lid.py", "atlas.py", "mathset.py", "grammar.py", "replication.py", "analyze.py", "build.py",
"PREREGISTRATION.md"):
shutil.copy(f, f"{D}/{f}")
os.makedirs(f"{D}/code", exist_ok=True)
for f in ("beam.py", "lid.py", "atlas.py", "mathset.py", "grammar.py", "replication.py", "analyze.py", "build.py"):
shutil.move(f"{D}/{f}", f"{D}/code/{f}")
for f in ("benak_lang.py", "benak_reward.py", "thinking.py", "relid.py", "routecheck.py", "distill.py"):
shutil.copy(f, f"{D}/code/{f}")
shutil.copy("data/atlas/templates.json", f"{D}/code/prompt_templates.json") # benak_lang finds it next to itself
if os.path.exists("CARD.md"): # live counts so the card never goes stale
card = open("CARD.md", encoding="utf-8").read()
card = card.replace("| `sea_thinking` | see file |", f"| `sea_thinking` | {len(think):,} |")
card = card.replace("| `atlas_runs` | 3,040 |", f"| `atlas_runs` | {len(atlas):,} |").replace("| `math_runs` | 903 |", f"| `math_runs` | {len(math_):,} |")
open(f"{D}/README.md", "w", encoding="utf-8").write(card)
print(f"atlas {len(atlas)} · math {len(math_)} · sea_thinking {len(think)} · cost ${R['cost_usd']}")
if a.push:
from huggingface_hub import HfApi
api = HfApi()
api.upload_folder(folder_path=D, repo_id="PugarHuda/benak", repo_type="dataset", commit_message="Benak release")
api.create_repo("PugarHuda/benak", repo_type="space", space_sdk="static", exist_ok=True, private=True)
api.upload_folder(folder_path=S, repo_id="PugarHuda/benak", repo_type="space", commit_message="Benak space")
print("pushed")
|