PugarHuda commited on
Commit
7debf49
·
verified ·
1 Parent(s): 96ac6f7

Benak release

Browse files
README.md CHANGED
@@ -55,9 +55,9 @@ configs:
55
 
56
  | config | rows | what |
57
  |---|---|---|
58
- | `atlas_runs` | 3,040 | Belebele, 19 languages × 40 parallel items × 4 prompting conditions (+ effort=high subset). Every call is included, with full reasoning, answer, grade, token usage and GlotLID profile. |
59
- | `math_runs` | 903 | 58 code-verified math word problems × 9 languages × {C0, C2}. |
60
- | `sea_thinking` | see file | **The fix.** Correct answers whose reasoning is ≥ 80% in the target language, in the [HuggingFaceH4/Multilingual-Thinking](https://huggingface.co/datasets/HuggingFaceH4/Multilingual-Thinking) schema (`reasoning_language, developer, user, analysis, final, messages`). That is the format Reflection's own TRL recipe uses, so it can be used directly for SFT. `source` marks native traces vs Beam-rewritten traces. |
61
  | `replication_profiled` | 4,139 | Beam traces that other entrants had already published (neulketing Korean, esscor TamilBench), re-scored with our detector. These cost $0 of inference. |
62
  | `grammar_book_bugis.json` | 20 calls | Can a book in Beam's 1M-token context teach it Buginese? Includes hypotheses, chrF++ scores and the reasoning-language probe. |
63
 
@@ -78,14 +78,39 @@ configs:
78
  * **H1** PASS: 15/18 languages with C0 eng_share >= 0.5
79
  * **H2** PASS: C2 > C0 in 16/18; of those 15 still < 0.8
80
  * **H3** PASS: C3 > C2 in 13/18
81
- * **H4** PASS: Spearman(Joshi tier, C2 target_share) = 0.281 over 18 languages (pre-registration phrased it as a negative correlation with 'low-resource-ness'; tier is ordered high=5)
82
- * **H5** PASS: mean acc(C2)-acc(C0) = 0.12 pts over 9 langs
83
- * **H6** FAIL: x2ind: [33.64, 26.99, 33.71, 49.12]; ind2x: [23.88, 24.21, 23.7, 32.73]
84
- * **H7** PASS: K3 think target_share = 0.438
 
 
 
 
 
 
 
 
 
 
 
85
 
86
  ## Grammar book → Buginese (MTOB-style)
87
 
88
- chrF++ on 100 NusaX sentences, buginese→indonesian / indonesian→buginese: no context 33.64/23.88; dictionary only 26.99/24.21; grammar only 33.71/23.7; **grammar+dictionary+582 example pairs (403k tokens in context) 49.12/32.73**. The dictionary alone hurts, so H6 fails as pre-registered.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
89
 
90
  ## How this relates to other entries
91
  * **[TamilBench](https://huggingface.co/datasets/esscor/TamilBench)** (esscor) and **[Korean Bench](https://huggingface.co/datasets/neulketing/beam-korean-grammar-mini)** (neulketing) published Beam traces. We re-read all 4,139 of them with the same detector (`replication_profiled`); see the Space for the table.
 
55
 
56
  | config | rows | what |
57
  |---|---|---|
58
+ | `atlas_runs` | 3,200 | Belebele, 19 languages × 40 parallel items × 4 prompting conditions (+ effort=high subset). Every call is included, with full reasoning, answer, grade, token usage and GlotLID profile. |
59
+ | `math_runs` | 1,276 | 58 code-verified math word problems × 9 languages × {C0, C2}, plus C3 (the `benak_lang` wrapper) for Indonesian, Javanese, Sundanese and Thai. |
60
+ | `sea_thinking` | 1,453 | **The fix.** Correct answers whose reasoning is ≥ 80% in the target language, in the [HuggingFaceH4/Multilingual-Thinking](https://huggingface.co/datasets/HuggingFaceH4/Multilingual-Thinking) schema (`reasoning_language, developer, user, analysis, final, messages`). That is the format Reflection's own TRL recipe uses, so it can be used directly for SFT. `source` marks native traces vs Beam-rewritten traces. |
61
  | `replication_profiled` | 4,139 | Beam traces that other entrants had already published (neulketing Korean, esscor TamilBench), re-scored with our detector. These cost $0 of inference. |
62
  | `grammar_book_bugis.json` | 20 calls | Can a book in Beam's 1M-token context teach it Buginese? Includes hypotheses, chrF++ scores and the reasoning-language probe. |
63
 
 
78
  * **H1** PASS: 15/18 languages with C0 eng_share >= 0.5
79
  * **H2** PASS: C2 > C0 in 16/18; of those 15 still < 0.8
80
  * **H3** PASS: C3 > C2 in 13/18
81
+ * **H4** PASS, weak: Spearman(Joshi tier, C2 target_share) = 0.281 over 18 languages, permutation p = 0.27. The direction is as predicted but not significant. *Deviation:* the pre-registration named NLLB high/low labels. Those are binary, so we used the 6-level [Joshi et al. 2020 taxonomy](https://microsoft.github.io/linguisticdiversity/assets/lang2tax.txt) and state the sign explicitly (tier 5 = high resource).
82
+ * **H5** PASS: mean acc(C2) − acc(C0) = +0.12 pts over 9 languages. Paired exact McNemar on the same problems gives p = 1.0 (7 vs 6 discordant).
83
+ * **H6** FAIL: bug→ind [33.64, 26.99, 33.71, 49.12]; ind→bug [23.88, 24.21, 23.7, 32.73]. The dictionary alone *lowers* bug→ind.
84
+ * **H7** PASS *as pre-registered* (K3 target_share = 0.438), **but this is quoting**. Once segments that quote the Buginese input are removed, Beam's reasoning in Buginese is 0% (K0), 22% (K1), 25% (K2) and **7% (K3)**. With the book in context, Beam thinks **in Indonesian** (88%), not in Buginese and not in English.
85
+
86
+ ## Robustness: is "60% English" an artefact?
87
+
88
+ | condition | English share (pre-registered, GlotLID) | quote-stripped GlotLID | Lingua (2nd detector) | Lingua, quote-stripped |
89
+ |---|---|---|---|---|
90
+ | C0 | 60% | 81% | 59% | 77% |
91
+ | C1 | 57% | 76% | 55% | 71% |
92
+ | C2 | 42% | 60% | 41% | 55% |
93
+ | C3 | 34% | 46% | 33% | 42% |
94
+
95
+ GlotLID and [Lingua](https://github.com/pemistahl/lingua-py) agree on whether a trace is mostly English in 84% of 3,040 traces. Quoted sentences, where Beam copies the passage or options, are written in the target language, so removing them **raises** the English share. The pre-registered number is the conservative one.
96
 
97
  ## Grammar book → Buginese (MTOB-style)
98
 
99
+ chrF++ on 100 NusaX sentences (bug→ind / ind→bug): no context 33.64 / 23.88; dictionary only (249k tokens) 26.99 / 24.21; grammar only (15k) 33.71 / 23.70; **grammar + dictionary + 582 example pairs (403k tokens in context) 49.12 / 32.73**. With no context, Beam "translates" into Buginese mostly by copying Indonesian.
100
+
101
+ ## Use the fix: `benak_lang`
102
+
103
+ `code/benak_lang.py` (40 lines, with templates for 19 languages) works with any OpenAI-compatible client. Its default is **C2**, which is robust across tasks. `primer=True` gives C3:
104
+
105
+ ```python
106
+ from benak_lang import think_in
107
+ msgs = think_in([{"role": "user", "content": "Berapa 17 x 23?"}], "ind_Latn")
108
+ client.chat.completions.create(model="Beam-1.0-501B-A23B", messages=msgs, extra_body={"reasoning_effort": "low"})
109
+ ```
110
+
111
+ **Transfer test** (`math_runs`, cond C3, n = 58 per language). The C3 primer is a reading-comprehension item, and it does **not** transfer to math. English share of reasoning, C0 → C2 → C3: Indonesian 70 → **19** → 23%, Javanese 70 → 55 → **50**%, Sundanese 69 → **56** → 68%, Thai 81 → **51** → 67%. Accuracy is 93–97% everywhere. A replayed-reasoning primer helps only when it resembles the task, so the wrapper defaults to C2.
112
+
113
+ Related work: Qi et al., *When Models Reason in Your Language* (EMNLP Findings 2025, [arXiv 2505.22888](https://arxiv.org/abs/2505.22888)), found that forcing in-language thinking costs accuracy on distilled reasoning models. On Beam we find **no measurable accuracy cost** (Belebele 91.8 → 91.8% from C0 to C3; math McNemar p = 1.0).
114
 
115
  ## How this relates to other entries
116
  * **[TamilBench](https://huggingface.co/datasets/esscor/TamilBench)** (esscor) and **[Korean Bench](https://huggingface.co/datasets/neulketing/beam-korean-grammar-mini)** (neulketing) published Beam traces. We re-read all 4,139 of them with the same detector (`replication_profiled`); see the Space for the table.
code/analyze.py CHANGED
@@ -100,8 +100,23 @@ rho = spearman([t for t, _ in h4], [s for _, s in h4]) if len(h4) > 2 else None
100
  # H4 as written: lower-resource -> less compliance, i.e. compliance rises with tier -> rho(tier, share) > 0
101
  V["H4"] = {"pass": rho is not None and rho > 0, "detail": f"Spearman(Joshi tier, C2 target_share) = {rho} over {len(h4)} languages "
102
  "(pre-registration phrased it as a negative correlation with 'low-resource-ness'; tier is ordered high=5)"}
 
 
 
 
 
 
103
  R["tiers"] = tier
104
- R["verdicts"] = V
 
 
 
 
 
 
 
 
 
105
 
106
  # ---- exploratory / robustness ----
107
  import random
 
100
  # H4 as written: lower-resource -> less compliance, i.e. compliance rises with tier -> rho(tier, share) > 0
101
  V["H4"] = {"pass": rho is not None and rho > 0, "detail": f"Spearman(Joshi tier, C2 target_share) = {rho} over {len(h4)} languages "
102
  "(pre-registration phrased it as a negative correlation with 'low-resource-ness'; tier is ordered high=5)"}
103
+ if rho is not None: # exact-ish permutation p-value (two-sided) for the weak correlation
104
+ prng = __import__("random").Random(1); ts_ = [t for t, _ in h4]; ss = [s for _, s in h4]; hits = 0
105
+ for _ in range(5000):
106
+ prng.shuffle(ss)
107
+ hits += abs(spearman(ts_, ss) or 0) >= abs(rho)
108
+ V["H4"]["detail"] += f"; permutation p = {hits / 5000:.3f} (direction as predicted, not significant)"
109
  R["tiers"] = tier
110
+ # Grammar 'think in Buginese' probe, quote-stripped: the x2ind probe quotes Buginese input sentences in its reasoning
111
+ if "grammar" in R:
112
+ from lid import profile as _prof
113
+ gtest = [json.loads(l) for l in open("data/bugis/test.jsonl", encoding="utf-8")]
114
+ for r in json.load(open("data/bugis/results.json", encoding="utf-8"))["rows"]:
115
+ if r["think"]:
116
+ src = "\n".join(s["bug"] for s in gtest[r["batch"] * 50:(r["batch"] + 1) * 50])
117
+ R["grammar"][f"{r['cond']}/think_noquote"] = _prof(r["reasoning"], "bug_Latn", exclude=src)
118
+ R["verdicts"] = dict(sorted(V.items()))
119
+ V = R["verdicts"]
120
 
121
  # ---- exploratory / robustness ----
122
  import random
code/benak_lang.py ADDED
@@ -0,0 +1,57 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """benak_lang: make Beam *think* in the user's language.
2
+
3
+ Default (primer=False) = Benak condition C2: "reason in X" as a system prompt. Robust across tasks: English share of
4
+ Beam's reasoning 60% -> 42% on Belebele (18 languages) and 70% -> 19% (Indonesian) on code-verified math, accuracy unchanged.
5
+ primer=True = C3: additionally replays one earlier turn whose `reasoning_content` is written in X. Best on reading
6
+ comprehension (-> 34%), but its reading-item primer does NOT transfer to math (Sundanese 56% -> 68%, Thai 51% -> 67%),
7
+ so only use it when your task resembles the primer. Templates for 19 languages ship in prompt_templates.json.
8
+
9
+ from openai import OpenAI
10
+ from benak_lang import think_in
11
+ client = OpenAI(base_url="https://api.reflection.ai/openai/v1", api_key=...)
12
+ msgs = think_in([{"role": "user", "content": "Berapa 17 x 23?"}], "ind_Latn")
13
+ client.chat.completions.create(model="Beam-1.0-501B-A23B", messages=msgs, extra_body={"reasoning_effort": "low"})
14
+ """
15
+ import json, os
16
+
17
+ _T = None
18
+ _EX_Q = {"role": "user", "content": None}
19
+
20
+
21
+ def templates(path=None):
22
+ global _T
23
+ if _T is None:
24
+ here = os.path.dirname(os.path.abspath(__file__))
25
+ for p in (path, os.path.join(here, "prompt_templates.json"), os.path.join(here, "data", "atlas", "templates.json"),
26
+ os.path.join(here, "..", "data", "prompt_templates.json")):
27
+ if p and os.path.exists(p):
28
+ _T = json.load(open(p, encoding="utf-8"))
29
+ break
30
+ else:
31
+ raise FileNotFoundError("prompt_templates.json not found")
32
+ return _T
33
+
34
+
35
+ def think_in(messages, lang, primer=False):
36
+ """Return a new message list that asks Beam to reason in `lang` (FLORES code, e.g. 'jav_Latn').
37
+ primer=False -> C2 (system prompt only, robust default); True -> C3 (adds the replayed-reasoning primer turn)."""
38
+ t = templates()[lang]
39
+ sys_text = t["reason_instruction"]
40
+ rest = [m for m in messages if m["role"] != "system"]
41
+ sys_msgs = [m for m in messages if m["role"] == "system"]
42
+ system = {"role": "system", "content": (sys_msgs[0]["content"] + "\n\n" if sys_msgs else "") + sys_text}
43
+ if not primer:
44
+ return [system, *rest]
45
+ # exactly the Benak C3 primer: a Belebele item in X and Beam-style reasoning about it, written in X
46
+ ex = [{"role": "user", "content": t["primer_user"]},
47
+ {"role": "assistant", "content": t["primer_answer"], "reasoning_content": t["example_reasoning"]}]
48
+ return [system, *ex, *rest]
49
+
50
+
51
+ if __name__ == "__main__":
52
+ m = think_in([{"role": "system", "content": "Be brief."}, {"role": "user", "content": "Berapa 17 x 23?"}], "ind_Latn",
53
+ primer=True)
54
+ assert m[0]["role"] == "system" and "Be brief." in m[0]["content"] and "bahasa Indonesia" in m[0]["content"]
55
+ assert m[2]["reasoning_content"] and m[-1]["content"] == "Berapa 17 x 23?" and len(m) == 4
56
+ assert len(think_in([{"role": "user", "content": "x"}], "jav_Latn")) == 2 # default = C2
57
+ print("benak_lang self-check ok")
code/build.py CHANGED
@@ -1,108 +1,121 @@
1
- """Assemble the release: HF dataset repo (release/dataset) + static Space (release/space).
2
- Usage: python build.py # build only
3
- python build.py --push # build + upload (dataset PugarHuda/benak, space PugarHuda/benak)"""
4
- import argparse, json, os, random, re, shutil
5
- from lid import label, is_target
6
-
7
- ap = argparse.ArgumentParser(); ap.add_argument("--push", action="store_true"); a = ap.parse_args()
8
- D, S = "release/dataset", "release/space"
9
- for p in (D, S):
10
- shutil.rmtree(p, ignore_errors=True)
11
- os.makedirs(f"{D}/data", exist_ok=True); os.makedirs(S, exist_ok=True)
12
- load = lambda p: [json.loads(l) for l in open(p, encoding="utf-8")] if os.path.exists(p) else []
13
- NAMES = json.load(open("data/atlas/templates.json", encoding="utf-8"))
14
-
15
- # 1) raw runs (every call, failures included) -------------------------------------------------------------
16
- atlas = load("data/atlas/runs_low.jsonl") + load("data/atlas/runs_high.jsonl")
17
- math_ = load("data/math/runs_low.jsonl") + load("data/math/runs_high.jsonl")
18
- for name, rows in (("atlas_runs", atlas), ("math_runs", math_)):
19
- with open(f"{D}/data/{name}.jsonl", "w", encoding="utf-8") as f:
20
- f.writelines(json.dumps(r, ensure_ascii=False) + "\n" for r in rows)
21
- shutil.copy("data/math/problems.json", f"{D}/data/math_problems.json")
22
- shutil.copy("data/math/parallel.json", f"{D}/data/math_parallel.json")
23
- shutil.copy("data/replication/profiled.jsonl", f"{D}/data/replication_profiled.jsonl")
24
- shutil.copy("data/lid_calibration.json", f"{D}/data/lid_calibration.json")
25
- shutil.copy("data/atlas/templates.json", f"{D}/data/prompt_templates.json")
26
- if os.path.exists("data/bugis/results.json"): # model outputs + scores only; the CC BY-NC dictionary is NOT redistributed
27
- shutil.copy("data/bugis/results.json", f"{D}/data/grammar_book_bugis.json")
28
- shutil.copy("data/bugis/test.jsonl", f"{D}/data/grammar_book_bugis_test.jsonl")
29
-
30
- # 2) sea_thinking: Multilingual-Thinking-format traces that are correct AND reasoned in the target language -----
31
- probs = {p["id"]: p for p in json.load(open("data/math/problems.json", encoding="utf-8"))}
32
- par = json.load(open("data/math/parallel.json", encoding="utf-8"))
33
- rew = load("data/thinking/rewritten.jsonl")
34
- think = []
35
- for r in math_ + atlas + rew:
36
- ts = (r.get("lid") or {}).get("target_share")
37
- if r.get("correct") and ts is not None and ts >= 0.8 and r.get("content"):
38
- lang = r["lang"]
39
- user = (par[str(r["id"])][lang]["text"] if "answer" in r and "id" in r and str(r["id"]) in par
40
- else r.get("user") or "")
41
- if not user:
42
- continue
43
- think.append({"reasoning_language": NAMES.get(lang, {}).get("lang_name_native", lang), "lang": lang,
44
- "developer": "", "user": user, "analysis": r["reasoning"], "final": r["content"],
45
- "messages": [{"role": "user", "content": user, "thinking": None},
46
- {"role": "assistant", "content": r["content"], "thinking": r["reasoning"]}],
47
- "source": r.get("trace_source", "native"), "task": "math" if "answer" in r else "belebele",
48
- "effort": r.get("effort"), "target_share": ts})
49
- with open(f"{D}/data/sea_thinking.jsonl", "w", encoding="utf-8") as f:
50
- f.writelines(json.dumps(t, ensure_ascii=False) + "\n" for t in think)
51
-
52
- # 3) results + space ------------------------------------------------------------------------------------------
53
- R = json.load(open("data/results.json", encoding="utf-8"))
54
- R["cost_usd"] = __import__("beam").spent()
55
- R["sea_thinking_rows"] = len(think)
56
- json.dump(R, open(f"{D}/results.json", "w", encoding="utf-8"), ensure_ascii=False, indent=1)
57
- json.dump(R, open(f"{S}/results.json", "w", encoding="utf-8"), ensure_ascii=False)
58
-
59
-
60
- def segs(text, lang):
61
- out = []
62
- for piece in re.split(r"(?<=[.!?。])\s+|\n+", text or ""):
63
- p = piece.strip()
64
- if not p:
65
- continue
66
- letters = sum(c.isalpha() for c in p)
67
- if letters >= 12:
68
- lab = label(p)[0]
69
- kind = "tgt" if is_target(lab, lang) else "eng" if lab == "eng_Latn" else "oth"
70
- else:
71
- kind = "tgt" # math / short fragments: neutral styling
72
- out.append([p, kind])
73
- return out
74
-
75
-
76
- samples = {}
77
- rng = random.Random(1)
78
- for r in [x for x in atlas if x["effort"] == "low"]:
79
- samples.setdefault(r["lang"], {}).setdefault(r["cond"], []).append(r)
80
- for l in samples:
81
- for c in samples[l]:
82
- pick = rng.sample(samples[l][c], min(4, len(samples[l][c])))
83
- samples[l][c] = [{"segs": segs(x["reasoning"], l), "correct": x["correct"],
84
- "eng_share": x["lid"]["eng_share"], "target_share": x["lid"]["target_share"]} for x in pick]
85
- json.dump(samples, open(f"{S}/samples.json", "w", encoding="utf-8"), ensure_ascii=False)
86
- shutil.copy("space/index.html", f"{S}/index.html")
87
- open(f"{S}/README.md", "w", encoding="utf-8").write(
88
- "---\ntitle: Benak\nemoji: 🧠\ncolorFrom: blue\ncolorTo: gray\nsdk: static\npinned: false\nlicense: cc-by-sa-4.0\n"
89
- "short_description: What language does Beam think in? 19 languages, 4 prompts\n---\n")
90
-
91
- # 4) code + docs ----------------------------------------------------------------------------------------------
92
- for f in ("beam.py", "lid.py", "atlas.py", "mathset.py", "grammar.py", "replication.py", "analyze.py", "build.py",
93
- "PREREGISTRATION.md"):
94
- shutil.copy(f, f"{D}/{f}")
95
- os.makedirs(f"{D}/code", exist_ok=True)
96
- for f in ("beam.py", "lid.py", "atlas.py", "mathset.py", "grammar.py", "replication.py", "analyze.py", "build.py"):
97
- shutil.move(f"{D}/{f}", f"{D}/code/{f}")
98
- if os.path.exists("CARD.md"):
99
- shutil.copy("CARD.md", f"{D}/README.md")
100
- print(f"atlas {len(atlas)} · math {len(math_)} · sea_thinking {len(think)} · cost ${R['cost_usd']}")
101
-
102
- if a.push:
103
- from huggingface_hub import HfApi
104
- api = HfApi()
105
- api.upload_folder(folder_path=D, repo_id="PugarHuda/benak", repo_type="dataset", commit_message="Benak release")
106
- api.create_repo("PugarHuda/benak", repo_type="space", space_sdk="static", exist_ok=True, private=True)
107
- api.upload_folder(folder_path=S, repo_id="PugarHuda/benak", repo_type="space", commit_message="Benak space")
108
- print("pushed")
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Assemble the release: HF dataset repo (release/dataset) + static Space (release/space).
2
+ Usage: python build.py # build only
3
+ python build.py --push # build + upload (dataset PugarHuda/benak, space PugarHuda/benak)"""
4
+ import argparse, json, os, random, re, shutil
5
+ from lid import label, is_target
6
+
7
+ ap = argparse.ArgumentParser(); ap.add_argument("--push", action="store_true"); a = ap.parse_args()
8
+ D, S = "release/dataset", "release/space"
9
+ for p in (D, S):
10
+ shutil.rmtree(p, ignore_errors=True)
11
+ os.makedirs(f"{D}/data", exist_ok=True); os.makedirs(S, exist_ok=True)
12
+ load = lambda p: [json.loads(l) for l in open(p, encoding="utf-8")] if os.path.exists(p) else []
13
+ NAMES = json.load(open("data/atlas/templates.json", encoding="utf-8"))
14
+
15
+ # 1) raw runs (every call, failures included) -------------------------------------------------------------
16
+ atlas = load("data/atlas/runs_low.jsonl") + load("data/atlas/runs_high.jsonl")
17
+ math_ = load("data/math/runs_low.jsonl") + load("data/math/runs_high.jsonl")
18
+ for name, rows in (("atlas_runs", atlas), ("math_runs", math_)):
19
+ with open(f"{D}/data/{name}.jsonl", "w", encoding="utf-8") as f:
20
+ f.writelines(json.dumps(r, ensure_ascii=False) + "\n" for r in rows)
21
+ shutil.copy("data/math/problems.json", f"{D}/data/math_problems.json")
22
+ shutil.copy("data/math/parallel.json", f"{D}/data/math_parallel.json")
23
+ shutil.copy("data/replication/profiled.jsonl", f"{D}/data/replication_profiled.jsonl")
24
+ shutil.copy("data/lid_calibration.json", f"{D}/data/lid_calibration.json")
25
+ shutil.copy("data/atlas/templates.json", f"{D}/data/prompt_templates.json")
26
+ if os.path.exists("data/bugis/results.json"): # model outputs + scores only; the CC BY-NC dictionary is NOT redistributed
27
+ shutil.copy("data/bugis/results.json", f"{D}/data/grammar_book_bugis.json")
28
+ shutil.copy("data/bugis/test.jsonl", f"{D}/data/grammar_book_bugis_test.jsonl")
29
+
30
+ # 2) sea_thinking: Multilingual-Thinking-format traces that are correct AND reasoned in the target language -----
31
+ probs = {p["id"]: p for p in json.load(open("data/math/problems.json", encoding="utf-8"))}
32
+ par = json.load(open("data/math/parallel.json", encoding="utf-8"))
33
+ rew = load("data/thinking/rewritten.jsonl")
34
+ think = []
35
+ _bel = {}
36
+ for r in math_ + atlas + rew:
37
+ ts = (r.get("lid") or {}).get("target_share")
38
+ if r.get("correct") and ts is not None and ts >= 0.8 and r.get("content"):
39
+ lang = r["lang"]
40
+ if "link" in r: # Belebele item: rebuild the exact prompt Beam saw (C0-C3 share the same user turn)
41
+ from atlas import mc_prompt
42
+ if lang not in _bel:
43
+ from datasets import load_dataset
44
+ _bel[lang] = {(x["link"], x["question_number"]): x for x in load_dataset("facebook/belebele", lang, split="test")}
45
+ user = mc_prompt(_bel[lang][(r["link"], r["qnum"])], NAMES[lang])
46
+ else:
47
+ user = r.get("user") or (par[str(r["id"])][lang]["text"] if str(r.get("id")) in par else "")
48
+ if not user:
49
+ continue
50
+ think.append({"reasoning_language": NAMES.get(lang, {}).get("lang_name_native", lang), "lang": lang,
51
+ "developer": "", "user": user, "analysis": r["reasoning"], "final": r["content"],
52
+ "messages": [{"role": "user", "content": user, "thinking": None},
53
+ {"role": "assistant", "content": r["content"], "thinking": r["reasoning"]}],
54
+ "source": r.get("trace_source", "native"), "task": "math" if "answer" in r else "belebele",
55
+ "effort": r.get("effort"), "target_share": ts})
56
+ with open(f"{D}/data/sea_thinking.jsonl", "w", encoding="utf-8") as f:
57
+ f.writelines(json.dumps(t, ensure_ascii=False) + "\n" for t in think)
58
+
59
+ # 3) results + space ------------------------------------------------------------------------------------------
60
+ R = json.load(open("data/results.json", encoding="utf-8"))
61
+ R["cost_usd"] = __import__("beam").spent()
62
+ R["sea_thinking_rows"] = len(think)
63
+ json.dump(R, open(f"{D}/results.json", "w", encoding="utf-8"), ensure_ascii=False, indent=1)
64
+ json.dump(R, open(f"{S}/results.json", "w", encoding="utf-8"), ensure_ascii=False)
65
+
66
+
67
+ def segs(text, lang):
68
+ out = []
69
+ for piece in re.split(r"(?<=[.!?。])\s+|\n+", text or ""):
70
+ p = piece.strip()
71
+ if not p:
72
+ continue
73
+ letters = sum(c.isalpha() for c in p)
74
+ if letters >= 12:
75
+ lab = label(p)[0]
76
+ kind = "tgt" if is_target(lab, lang) else "eng" if lab == "eng_Latn" else "oth"
77
+ else:
78
+ kind = "tgt" # math / short fragments: neutral styling
79
+ out.append([p, kind])
80
+ return out
81
+
82
+
83
+ samples = {}
84
+ rng = random.Random(1)
85
+ for r in [x for x in atlas if x["effort"] == "low"]:
86
+ samples.setdefault(r["lang"], {}).setdefault(r["cond"], []).append(r)
87
+ for l in samples:
88
+ for c in samples[l]:
89
+ pick = rng.sample(samples[l][c], min(4, len(samples[l][c])))
90
+ samples[l][c] = [{"segs": segs(x["reasoning"], l), "correct": x["correct"],
91
+ "eng_share": x["lid"]["eng_share"], "target_share": x["lid"]["target_share"]} for x in pick]
92
+ json.dump(samples, open(f"{S}/samples.json", "w", encoding="utf-8"), ensure_ascii=False)
93
+ shutil.copy("space/index.html", f"{S}/index.html")
94
+ open(f"{S}/README.md", "w", encoding="utf-8").write(
95
+ "---\ntitle: Benak\nemoji: 🧠\ncolorFrom: blue\ncolorTo: gray\nsdk: static\npinned: false\nlicense: cc-by-sa-4.0\n"
96
+ "short_description: What language does Beam think in? 19 languages, 4 prompts\n---\n")
97
+
98
+ # 4) code + docs ----------------------------------------------------------------------------------------------
99
+ for f in ("beam.py", "lid.py", "atlas.py", "mathset.py", "grammar.py", "replication.py", "analyze.py", "build.py",
100
+ "PREREGISTRATION.md"):
101
+ shutil.copy(f, f"{D}/{f}")
102
+ os.makedirs(f"{D}/code", exist_ok=True)
103
+ for f in ("beam.py", "lid.py", "atlas.py", "mathset.py", "grammar.py", "replication.py", "analyze.py", "build.py"):
104
+ shutil.move(f"{D}/{f}", f"{D}/code/{f}")
105
+ for f in ("benak_lang.py", "thinking.py", "relid.py"):
106
+ shutil.copy(f, f"{D}/code/{f}")
107
+ shutil.copy("data/atlas/templates.json", f"{D}/code/prompt_templates.json") # benak_lang finds it next to itself
108
+ if os.path.exists("CARD.md"): # live counts so the card never goes stale
109
+ card = open("CARD.md", encoding="utf-8").read()
110
+ card = card.replace("| `sea_thinking` | see file |", f"| `sea_thinking` | {len(think):,} |")
111
+ card = card.replace("| `atlas_runs` | 3,040 |", f"| `atlas_runs` | {len(atlas):,} |").replace("| `math_runs` | 903 |", f"| `math_runs` | {len(math_):,} |")
112
+ open(f"{D}/README.md", "w", encoding="utf-8").write(card)
113
+ print(f"atlas {len(atlas)} · math {len(math_)} · sea_thinking {len(think)} · cost ${R['cost_usd']}")
114
+
115
+ if a.push:
116
+ from huggingface_hub import HfApi
117
+ api = HfApi()
118
+ api.upload_folder(folder_path=D, repo_id="PugarHuda/benak", repo_type="dataset", commit_message="Benak release")
119
+ api.create_repo("PugarHuda/benak", repo_type="space", space_sdk="static", exist_ok=True, private=True)
120
+ api.upload_folder(folder_path=S, repo_id="PugarHuda/benak", repo_type="space", commit_message="Benak space")
121
+ print("pushed")
code/mathset.py CHANGED
@@ -161,6 +161,9 @@ def cmd_solve(a):
161
  msgs = [{"role": "user", "content": par[i][c]["text"]}]
162
  if cond == "C2":
163
  msgs.insert(0, {"role": "system", "content": SYS.get(c) or tps[c]["reason_instruction"]})
 
 
 
164
  r = chat(msgs, a.effort, budget, tag="math")
165
  b = boxed(r["content"])
166
  return {"id": i, "lang": c, "cond": cond, "effort": a.effort, "answer": probs[i]["answer"], "boxed": b,
@@ -170,8 +173,15 @@ def cmd_solve(a):
170
  "topic": probs[i]["topic"], "difficulty": probs[i]["difficulty"], "route": r.get("route"), "served_model": r.get("model"), "latency_s": r.get("latency_s")}
171
 
172
  rows = [res for _, res in run_many(one, jobs, 16, f"math-{a.effort}") if not isinstance(res, Exception)]
173
- with open(f"{OUT}/runs_{a.effort}.jsonl", "w", encoding="utf-8") as f:
174
- f.writelines(json.dumps(r, ensure_ascii=False) + "\n" for r in rows)
 
 
 
 
 
 
 
175
  print(f"{len(rows)}/{len(jobs)} solved-rows written; spent ${spent()}")
176
 
177
 
 
161
  msgs = [{"role": "user", "content": par[i][c]["text"]}]
162
  if cond == "C2":
163
  msgs.insert(0, {"role": "system", "content": SYS.get(c) or tps[c]["reason_instruction"]})
164
+ elif cond == "C3": # the shipped benak_lang wrapper; its primer is a reading item -> tests transfer to math
165
+ from benak_lang import think_in
166
+ msgs = think_in(msgs, c, primer=True)
167
  r = chat(msgs, a.effort, budget, tag="math")
168
  b = boxed(r["content"])
169
  return {"id": i, "lang": c, "cond": cond, "effort": a.effort, "answer": probs[i]["answer"], "boxed": b,
 
173
  "topic": probs[i]["topic"], "difficulty": probs[i]["difficulty"], "route": r.get("route"), "served_model": r.get("model"), "latency_s": r.get("latency_s")}
174
 
175
  rows = [res for _, res in run_many(one, jobs, 16, f"math-{a.effort}") if not isinstance(res, Exception)]
176
+ path = f"{OUT}/runs_{a.effort}.jsonl"
177
+ merged = {}
178
+ if os.path.exists(path): # merge: a run with other --conds/--langs must not erase earlier rows
179
+ for l in open(path, encoding="utf-8"):
180
+ d = json.loads(l); merged[(d["id"], d["lang"], d["cond"])] = d
181
+ for r in rows:
182
+ merged[(r["id"], r["lang"], r["cond"])] = r
183
+ with open(path, "w", encoding="utf-8") as f:
184
+ f.writelines(json.dumps(r, ensure_ascii=False) + "\n" for r in merged.values())
185
  print(f"{len(rows)}/{len(jobs)} solved-rows written; spent ${spent()}")
186
 
187
 
code/prompt_templates.json ADDED
@@ -0,0 +1,328 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "eng_Latn": {
3
+ "lang_name_native": "English",
4
+ "question_label": "Question",
5
+ "answer_instruction": "Answer with the number of the correct option (1-4) inside \\boxed{}.",
6
+ "reason_instruction": "Think and reason step by step in English.",
7
+ "example_reasoning": "The question asks which tip is NOT accurate. The passage says to keep the hand relaxed and to avoid extra finger motion, so options 2 and 3 are accurate tips. It says that on the accordion extra volume comes from using the bellows with more pressure or speed, so option 4 is also accurate. It explicitly says there is no need to hit the keys with a lot of force for extra volume, so option 1 is the inaccurate tip. The answer is 1.",
8
+ "source": "human",
9
+ "primer_user": "Make sure your hand is as relaxed as possible while still hitting all the notes correctly - also try not to make much extraneous motion with your fingers. This way, you will tire yourself out as little as possible. Remember there's no need to hit the keys with a lot of force for extra volume like on the piano. On the accordion, to get extra volume, you use the bellows with more pressure or speed.\n\nQuestion: According to the passage, what would not be considered an accurate tip for successfully playing the accordion?\n1. For additional volume, increase the force with which you hit the keys\n2. Keep unnecessary movement to a minimum in order to preserve your stamina\n3. Be mindful of hitting the notes while maintaining a relaxed hand\n4. Increase the speed with which you operate the bellows to achieve extra volume\n\nAnswer with the number of the correct option (1-4) inside \\boxed{}.",
10
+ "primer_answer": "\boxed{1}"
11
+ },
12
+ "ind_Latn": {
13
+ "lang_name_native": "bahasa Indonesia",
14
+ "question_label": "Pertanyaan",
15
+ "answer_instruction": "Jawab dengan nomor pilihan yang benar (1-4) di dalam \\boxed{}.",
16
+ "reason_instruction": "Berpikir dan bernalarlah langkah demi langkah dalam bahasa Indonesia.",
17
+ "example_reasoning": "Pertanyaannya menanyakan tips mana yang TIDAK tepat. Bacaan mengatakan tangan harus tetap rileks dan jari tidak banyak bergerak, jadi pilihan 2 dan 3 adalah tips yang tepat. Bacaan juga mengatakan bahwa pada akordeon volume tambahan didapat dengan menekan atau menggerakkan bellow lebih kuat atau lebih cepat, jadi pilihan 4 juga tepat. Bacaan dengan jelas mengatakan tidak perlu menekan tuts dengan keras untuk menambah volume, jadi pilihan 1 adalah tips yang tidak tepat. Jawabannya 1.",
18
+ "source": "human",
19
+ "primer_user": "Pastikan tangan dalam keadaan serileks mungkin sambil tetap menekan setiap nada dengan benar - Upayakan jari tidak membuat banyak gerakan ekstra. Dengan cara ini, Anda akan mengurangi rasa lelah Anda. Ingatlah bahwa tidak perlu menekan tuts terlalu keras untuk mendapatkan volume ekstra seperti pada piano. Di akordion, untuk mendapatkan volume lebih besar, Anda menggunakan ububan dengan tekanan atau kecepatan lebih besar.\n\nPertanyaan: Menurut kutipan tersebut, apa yang bukan merupakan tips akurat untuk memainkan akordion dengan sukses?\n1. Untuk volume yang lebih keras, tingkatkan kekuatan tekanan yang Anda gunakan untuk menekan tuts\n2. Buat seminimal mungkin gerakan yang tidak diperlukan untuk menjaga stamina Anda\n3. Perhatikan saat menekan nada sambil mempertahankan tangan yang rileks\n4. Tingkatkan kecepatan Anda dalam mengoperasikan ububan untuk suara yang lebih keras\n\nJawab dengan nomor pilihan yang benar (1-4) di dalam \\boxed{}.",
20
+ "primer_answer": "\boxed{1}"
21
+ },
22
+ "jav_Latn": {
23
+ "lang_name_native": "Basa Jawa",
24
+ "question_label": "Pitakonan",
25
+ "answer_instruction": "Wangsulana kanthi nomer pilihan sing bener (1-4) ing njero \\boxed{}.",
26
+ "reason_instruction": "Mikir lan nalar kanthi langkah demi langkah ing Basa Jawa.",
27
+ "example_reasoning": "Pitakonan iki takon saran sing ORA akurat. Teks kasebut ngandharake supaya tangan tetep santai lan ngindhari gerakan driji sing ekstra, mulane pilihan 2 lan 3 minangka saran sing akurat. Teks kasebut uga ngandharake yen ing akordeon volume ekstra asale saka panggunaan balg kanthi tekanan utawa kacepetan sing luwih, mulane pilihan 4 uga akurat. Kanthi tegas kasebut ora perlu mencet tombol kanthi kekuwatan gedhe kanggo volume ekstra, mulane pilihan 1 minangka saran sing ora akurat. Jawabane yaiku 1.",
28
+ "source": "beam-translated",
29
+ "lid_example": {
30
+ "n_seg": 5,
31
+ "target_share": 1.0,
32
+ "eng_share": 0.0,
33
+ "other_share": 0.0,
34
+ "first_eng_pos": null,
35
+ "top_other": null
36
+ },
37
+ "primer_user": "Mesthekake tanganmu anteng kambi isih nabuh kabeh calathu kanthi bener - uga coba driji-drijimu aja kakehan obah. Kanthi cara iki, sampeyan bakal ora krasa kesel banget. Kudu dieling-eling yen sampeyan ora kudu nggunakake tenaga sing gedhe kanggo mijet kunci kuwi kaya mijet piano. Ing akordian, kanggo entuk volume tambahan, sampeyan nggunakake embusan kanthi luwih mlenet lan cepet.\n\nPitakonan: Miturut pethikan kasebut, apa sing ora bisa dianggep tips ingkang bener kanggo dolan akordian?\n1. Kanggo volume tambahan, nggunakake tenaga sing gedhe kanggo mijet kunci\n2. Aja kakehan obah kanggo njaga stamina\n3. Mesthekake wektu nabuh calathu tangane anteng\n4. Nggunakake embusan kanthi luwih cepet kanggo entuk volume tambahan\n\nWangsulana kanthi nomer pilihan sing bener (1-4) ing njero \\boxed{}.",
38
+ "primer_answer": "\boxed{1}"
39
+ },
40
+ "sun_Latn": {
41
+ "lang_name_native": "Basa Sunda",
42
+ "question_label": "Patarosan",
43
+ "answer_instruction": "Jawab ku nomer pilihan anu bener (1-4) di jero \\boxed{}.",
44
+ "reason_instruction": "Pikirkeun jeung alesan hambalan demi hambalan dina Basa Sunda.",
45
+ "example_reasoning": "Patarosan nanyakeun tip mana anu TEU akurat. Dina tulisan disebutkeun sangkan leungeun tetep rileks jeung ulah aya gerakan ramo anu kaleuleuwihi, ku kituna pilihan 2 jeung 3 mangrupa tip anu akurat. Disebutkeun yén dina akordion volume tambahan asalna tina pamakean bellows kalayan tekanan atawa kacepetan anu leuwih, ku kituna pilihan 4 ogé akurat. Sacara écés disebutkeun yén teu perlu mencét kenop kalayan gaya anu kacida pikeun volume tambahan, ku kituna pilihan 1 mangrupa tip anu teu akurat. Jawabanna nyaéta 1.",
46
+ "source": "beam-translated",
47
+ "lid_example": {
48
+ "n_seg": 5,
49
+ "target_share": 1.0,
50
+ "eng_share": 0.0,
51
+ "other_share": 0.0,
52
+ "first_eng_pos": null,
53
+ "top_other": null
54
+ },
55
+ "primer_user": "Pastikeun leungeun anjeun sakumaha tiasa santai bari tetep pencét kabéh nada anu leres - ogé cobi ulah ngadamel seueur gerakan asing sareng ramo anjeun. Ku cara ieu, anjeun bakal capék padahal ngan maén sakedik. Émut, henteu kudu mencet konci sakuat tanaga pikeun nambah suara sapertos dina piano. Dinu akordion, pikeun ngengingkeun volume langkung, anggo bagian handap langkung teuas atanapi langkung enggal.\n\nPatarosan: Dumasar kana bagéan tina karangan anu disayogikeun, piwuruk mana anu teu luyu patali cara anu bener maénkeun akordion?\n1. Ngarah nambah tarik sorana, pencét konci anu diteken leuwih teuas\n2. Ulah seueur teuing gerak teu perlu pikeun ngajaga stamina anjeun\n3. Ati-ati dina mencét koncina bari tetep leungeun mah santai\n4. Paénkeun bagian lipetan akordion leuwih gancang ngarah volume langkung tarik deui\n\nJawab ku nomer pilihan anu bener (1-4) di jero \\boxed{}.",
56
+ "primer_answer": "\boxed{1}"
57
+ },
58
+ "zsm_Latn": {
59
+ "lang_name_native": "Bahasa Melayu",
60
+ "question_label": "Soalan",
61
+ "answer_instruction": "Jawab dengan nombor pilihan yang betul (1-4) di dalam \\boxed{}.",
62
+ "reason_instruction": "Berfikir dan berikan alasan langkah demi langkah dalam Bahasa Melayu.",
63
+ "example_reasoning": "Soalan ini menanya tip yang TIDAK tepat. Petikan itu menyatakan supaya tangan kekal santai dan mengelakkan gerakan jari yang berlebihan, jadi pilihan 2 dan 3 adalah tip yang tepat. Ia menyatakan bahawa pada akordion, kelantangan tambahan datang daripada penggunaan belos dengan tekanan atau kelajuan yang lebih tinggi, jadi pilihan 4 juga tepat. Ia menyatakan secara jelas bahawa tidak perlu menekan kekunci dengan kuat untuk mendapatkan kelantangan tambahan, jadi pilihan 1 adalah tip yang tidak tepat. Jawapannya ialah 1.",
64
+ "source": "beam-translated",
65
+ "lid_example": {
66
+ "n_seg": 5,
67
+ "target_share": 1.0,
68
+ "eng_share": 0.0,
69
+ "other_share": 0.0,
70
+ "first_eng_pos": null,
71
+ "top_other": null
72
+ },
73
+ "primer_user": "Pastikan tangan anda setenang yang boleh sementara masih menekan semua tanda dengan betul - selain itu, cuba untuk tidak membuat gerakan luaran dengan jari anda. Dengan cara ini, anda akan memenatkan diri sendiri sesedikit yang mungkin. Ingatlah bahawa tidak perlu menggunakan banyak tenaga untuk menekan kekunci untuk mendapat bunyi yang kuat seperti piano. Pada akordion, bagi mendapatkan kelantangan tambahan, anda guna pakai belos pada tekanan atau kelajuan yang lebih tinggi.\n\nSoalan: Menurut petikan, apakah petua yang dianggap tidak tepat untuk memainkan akordion dengan baik?\n1. Untuk menambah kelantangan, tingkatkan daya anda menekan kekunci\n2. Minimumkan pergerakan yang tidak diperlukan untuk menyimpan stamina anda\n3. Pastikan anda mencapai not sambil memastikan jari tenang\n4. Tingkatkan kelajuan anda mengendalikan belos untuk mendapatkan lebih kelantangan\n\nJawab dengan nombor pilihan yang betul (1-4) di dalam \\boxed{}.",
74
+ "primer_answer": "\boxed{1}"
75
+ },
76
+ "tgl_Latn": {
77
+ "lang_name_native": "Tagalog",
78
+ "question_label": "Tanong",
79
+ "answer_instruction": "Sagutin gamit ang numero ng tamang opsyon (1-4) sa loob ng \\boxed{}.",
80
+ "reason_instruction": "Mag-isip at magpaliwanag nang hakbang-hakbang sa Tagalog.",
81
+ "example_reasoning": "Tinatanong kung aling payo ang hindi tumpak. Sinasabi ng teksto na panatilihing relaxed ang kamay at iwasan ang sobrang paggalaw ng daliri, kaya ang mga opsyon 2 at 3 ay tumpak na payo. Sinasabi rin nito na sa akordiyon, ang dagdag na volume ay nagmumula sa paggamit ng bellows nang may mas mataas na presyon o bilis, kaya tumpak din ang opsyon 4. Malinaw nitong sinasabi na hindi kailangang pindutin nang malakas ang mga key para sa dagdag na volume, kaya ang opsyon 1 ang hindi tumpak na payo. Ang sagot ay 1.",
82
+ "source": "beam-translated",
83
+ "lid_example": {
84
+ "n_seg": 4,
85
+ "target_share": 1.0,
86
+ "eng_share": 0.0,
87
+ "other_share": 0.0,
88
+ "first_eng_pos": null,
89
+ "top_other": null
90
+ },
91
+ "primer_user": "Siguruhing kalmado ang iyong kamay hangga't posible habang wastong inaabot pa rin ang lahat ng mga nota - subukan ding huwag gumawa ng labis na paggalaw sa iyong mga daliri. Sa ganitong paraan, papagurin mo lamang ang sarili mo nang kaunti hangga't maaari. Tandaan, walang pangangailangan na pindutin nang madiin ang mga teklado para lumakas ang tunog katulad ng kapag sa piyano. Sa akordyon, upang makakuha ng karagdagang dami, gumamit ka ng mga bellow na mayroong higit na presyon at bilis.\n\nTanong: Ayon sa sipi, ano ang hindi maituturing na tumpak na paalala para sa matagumpay na pagpapatugtog ng akordyon?\n1. Para sa mas malakas na volume, dagdagan ang puwersa ng pagpindot sa teklado\n2. Hanggang maaari, iwasan ang hindi kinakailangang paggalaw upang mapreserba ang iyong stamina\n3. Maging maingat sa pag-abot ng nota habang nananatiling kalmado ang kamay\n4. Mas bilisan pa ang paggalaw ng mga bellow upang mas mapalakas ang volume\n\nSagutin gamit ang numero ng tamang opsyon (1-4) sa loob ng \\boxed{}.",
92
+ "primer_answer": "\boxed{1}"
93
+ },
94
+ "ceb_Latn": {
95
+ "lang_name_native": "Sinugbuanong Binisaya",
96
+ "question_label": "Pangutana",
97
+ "answer_instruction": "Tubaga sa numero sa hustong kapilian (1-4) sulod sa \\boxed{}.",
98
+ "reason_instruction": "Pagpamalandong ug paghinungdan hinay-hinay sa Sinugbuanong Binisaya.",
99
+ "example_reasoning": "Ang pangutana mangutana kung unsang tambag ang DILI tukma. Ang teksto nag-ingon nga kinahanglang ipabilin nga relaks ang kamot ug likayan ang sobrang paglihok sa tudlo, busa ang mga kapilian 2 ug 3 mao ang tukmang tambag. Nag-ingon usab kini nga sa akordiyon, ang dugang nga volume naggikan sa paggamit sa bellows nga adunay mas dakong presyur o mas paspas, busa ang kapilian 4 tukma usab. Tin-aw nga nag-ingon nga walay panginahanglan nga pug-on ang mga yawe nga adunay dakong kusog aron makadugang og volume, busa ang kapilian 1 mao ang dili tukmang tambag. Ang tubag mao ang 1.",
100
+ "source": "beam-translated",
101
+ "lid_example": {
102
+ "n_seg": 5,
103
+ "target_share": 1.0,
104
+ "eng_share": 0.0,
105
+ "other_share": 0.0,
106
+ "first_eng_pos": null,
107
+ "top_other": null
108
+ },
109
+ "primer_user": "Siguruha nga kalma ang imong kamot kutob sa mahimo samtang padayon nga saktong maabot ang tanang nota - paningkamuti usab nga dili mobuhat og way hinungdang lihok gamit ang imong mga tudlo. Sa kani nga pamaagi, gamay ra ang imong kahago kutob sa maarangan. Hinumdumi nga dili kinahanglanng ig-on ang mga tekla gamit ang grabeng puwersa para sa dugang kakusgon sama sa piano. Sa akordyon, aron makakuha og dugang nga kakusog, kinahanglan gamiton nimo ang mga hasohasan nga adunay labing presyur o tulin.\n\nPangutana: Sumala sa istorya, unsay dili makonsiderar ingong tukmang tip para sa malamposong pagpatukar og akordyon?\n1. Para dugangan ang kakusog, patas-i ang puwersa sa pag-igo nimo sa mga tekla\n2. Likayi ang wala gikinahanglang mga lihok aron mapreserbar ang imong stamina\n3. Bantayi ang pag-igo sa mga nota samtang nagmentinar og kalma nga kamot\n4. Patas-i ang katulin sa imong pag-operate sa mga hasohasan aron makab-ot ang dugang kakusog\n\nTubaga sa numero sa hustong kapilian (1-4) sulod sa \\boxed{}.",
110
+ "primer_answer": "\boxed{1}"
111
+ },
112
+ "vie_Latn": {
113
+ "lang_name_native": "Tiếng Việt",
114
+ "question_label": "Câu hỏi",
115
+ "answer_instruction": "Trả lời bằng số của lựa chọn đúng (1-4) trong \\boxed{}.",
116
+ "reason_instruction": "Hãy suy nghĩ và lập luận từng bước bằng Tiếng Việt.",
117
+ "example_reasoning": "Câu hỏi hỏi mẹo nào không chính xác. Đoạn văn nói nên giữ tay thư giãn và tránh chuyển động ngón tay thừa, nên các lựa chọn 2 và 3 là các mẹo chính xác. Nó nói rằng trên đàn accordion, âm lượng lớn hơn đến từ việc dùng phong cầm với áp lực hoặc tốc độ lớn hơn, nên lựa chọn 4 cũng chính xác. Nó nói rõ rằng không cần nhấn phím với lực lớn để có âm lượng lớn hơn, nên lựa chọn 1 là mẹo không chính xác. Câu trả lời là 1.",
118
+ "source": "beam-translated",
119
+ "lid_example": {
120
+ "n_seg": 4,
121
+ "target_share": 1.0,
122
+ "eng_share": 0.0,
123
+ "other_share": 0.0,
124
+ "first_eng_pos": null,
125
+ "top_other": null
126
+ },
127
+ "primer_user": "Hãy đảm bảo tay của bạn được thư giãn nhất có thể khi nhấn tất cả các nốt nhạc chính xác - và cố gắng đừng làm nhiều cử động ngón tay không cần thiết. Chơi đàn theo cách này sẽ giúp bạn giảm mất sức tối đa. Hãy nhớ rằng, bạn không cần phải dùng nhiều lực để ấn phím nhằm tăng âm lượng như khi chơi đàn piano. Trên đàn phong cầm, để tăng âm lượng, bạn sử dụng ống thổi tác đ��ng mạnh hơn hoặc nhanh hơn.\n\nCâu hỏi: Theo đoạn văn, đâu không được coi là lời khuyên đúng để chơi được đàn phong cầm?\n1. Để tăng âm lượng, dùng thêm nhiều lực khi ấn phím\n2. Hạn chế tối thiểu động tác thừa để giữ sức\n3. Chú ý khi nhấn các nốt nhạc phải giữ cho tay được thư giãn nhất có thể\n4. Tăng tốc độ di chuyển hộp xếp để tăng âm lượng\n\nTrả lời bằng số của lựa chọn đúng (1-4) trong \\boxed{}.",
128
+ "primer_answer": "\boxed{1}"
129
+ },
130
+ "tha_Thai": {
131
+ "lang_name_native": "ไทย",
132
+ "question_label": "คำถาม",
133
+ "answer_instruction": "ตอบด้วยหมายเลขของตัวเลือกที่ถูกต้อง (1-4) ภายใน \\boxed{}.",
134
+ "reason_instruction": "คิดและให้เหตุผลทีละขั้นตอนเป็นภาษาไทย",
135
+ "example_reasoning": "คำถามถามว่าเคล็ดลับข้อใดไม่ถูกต้อง บทความระบุให้ผ่อนคลายมือและหลีกเลี่ยงการขยับนิ้วเกินจำเป็น ดังนั้นตัวเลือก 2 และ 3 เป็นเคล็ดลับที่ถูกต้อง มันระบุว่าบนอคคอร์เดียน ความดังเพิ่มเติมมาจากการใช้ลูกเป่าด้วยแรงดันหรือความเร็วที่มากขึ้น ดังนั้นตัวเลือก 4 ก็ถูกต้องเช่นกัน มันระบุอย่างชัดเจนว่าไม่จำเป็นต้องกดคีย์ด้วยแรงมากเพื่อให้ดังขึ้น ดังนั้นตัวเลือก 1 เป็นเคล็ดลับที่ไม่ถูกต้อง คำตอบคือ 1",
136
+ "source": "beam-translated",
137
+ "lid_example": {
138
+ "n_seg": 1,
139
+ "target_share": 1.0,
140
+ "eng_share": 0.0,
141
+ "other_share": 0.0,
142
+ "first_eng_pos": null,
143
+ "top_other": null
144
+ },
145
+ "primer_user": "ปล่อยให้มือของคุณผ่อนคลายที่สุดเท่าที่จะทำได้ขณะที่ยังคงเล่นโน้ตทุกตัวได้อย่างถูกต้อง พยายามไม่เคลื่อนไหวนิ้วมือมากเกินไปโดยไม่จำเป็น วิธีนี้จะทำให้ลดความเหน็ดเหนื่อยที่คุณได้รับให้เหลือน้อยที่สุด จำไว้ว่าไม่จำเป็นต้องกดแป้นให้แรงเพื่อให้เสียงดังขึ้นเหมือนเปียโน หากต้องการเพิ่มเสียงของหีบเพลงให้ดังขึ้น ให้ใช้แรงดันหรือความเร็วของที่สูบลมให้มากขึ้น\n\nคำถาม: จากบทความ ข้อใดไม่ใช่เคล็ดลับที่ถูกต้องสำหรับการเล่นหีบเพลงให้ประสบความสำเร็จ\n1. หากต้องการให้เสียงดังขึ้น ให้เพิ่มแรงดันตอนเล่นตัวโน้ต\n2. พยายามไม่เคลื่อนไหวนิ้วมือมากเกินไปเพื่อไม่ให้ออกแรงมากเกินไป\n3. ระวังตอนเล่นตัวโน้ตขณะที่ปล่อยให้มือผ่อนคลายที่สุดไปด้วย\n4. เพิ่มความเร็วของที่สูบลมเพื่อทำให้เสียงดังขึ้น\n\nตอบด้วยหมายเลขของตัวเลือกที่ถูกต้อง (1-4) ภายใน \\boxed{}.",
146
+ "primer_answer": "\boxed{1}"
147
+ },
148
+ "khm_Khmr": {
149
+ "lang_name_native": "ខ្មែរ",
150
+ "question_label": "សំណួរ",
151
+ "answer_instruction": "ឆ្លើយដោយលេខនៃជម្រើសត្រឹមត្រូវ (1-4) នៅក្នុង \\boxed{}.",
152
+ "reason_instruction": "គិត និងពន្យល់ជំហានៗជាភាសាខ្មែរ.",
153
+ "example_reasoning": "សំណួរសួរថាតើគន្លឹះណាមួយមិនត្រឹមត្រូវ។ អត្ថបទបាននិយាយថាត្រូវរក្សាដៃឱ្យស្រួ�� និងជៀសវាងចលនាបន្ថែមរបស់ម្រាមដៃ ដូច្នេះគន្លឹះទី 2 និងទី 3 គឺត្រឹមត្រូវ។ វាបាននិយាយថានៅលើអាក័រដេអុង សំឡេងបន្ថែមមកពីការប្រើប៊ែលឡូសជាមួយសម្ពាធឬល្បឿនខ្លាំងជាង ដូច្នេះគន្លឹះទី 4 ក៏ត្រឹមត្រូវដែរ។ វាបាននិយាយយ៉ាងច្បាស់ថាមិនចាំបាច់វាយគ្រាប់ចុចដោយកម្លាំងច្រើនសម្រាប់សំឡេងបន្ថែម ដូច្នេះគន្លឹះទី 1 គឺជាគន្លឹះមិនត្រឹមត្រូវ។ ចម្លើយគឺ 1.",
154
+ "source": "beam-translated",
155
+ "lid_example": {
156
+ "n_seg": 1,
157
+ "target_share": 1.0,
158
+ "eng_share": 0.0,
159
+ "other_share": 0.0,
160
+ "first_eng_pos": null,
161
+ "top_other": null
162
+ },
163
+ "primer_user": "ត្រូវ​ឱ្យ​ប្រាកដ​ថា ដៃ​របស់​អ្នក​បាន​សម្រាក​តាម​ដែល​អាច​ធ្វើ​ទៅ​បាន នៅ​ខណៈ​ពេល​ដែល​អ្នក​អាច​នៅតែ​កត់ត្រា​បាន​ត្រឹមត្រូវ ម្យ៉ាងទៀត​ចូរ​កុំ​ព្យាយាម​ធ្វើ​ចលនា​ដោយ​ប្រើ​ម្រាម​ដៃ​របស់​អ្នក​បើ​មិន​ចាំបាច់។ វិធីនេះ​អ្នកនឹងធ្វើឱ្យខ្លួនឯងរាងហត់បន្តិច។ ចូរចងចាំថាមិន​ចាំបាច់​ចុច​តម្លើងសម្លេង​ដោយប្រើកម្លាំង​ខ្លាំង​ដូចនៅលើ​ព្យ៉ាណូទេ។ ដើម្បីបានកម្រិត​សំឡងពិសេសនៅលើអាគ័រឌីអុងនេះ ​អ្នកត្រូវប្រើប្រដាប់ផ្លុំខ្យល់​ដើម្បីបន្ថែម​សម្ពាធ ឬល្បឿន។\n\nសំណួរ: យោងតាមអត្ថបទសម្រង់នេះ តើអ្វីដែលមិនត្រូវបានចាត់ទុកថាជាគន្លឹះត្រឹមត្រូវសម្រាប់ការលេងអាគ័រដីយ៉ុងដោយជោគជ័យ?\n1. សម្រាប់កម្រិតសំឡេងបន្ថែម សូមបង្កើនកម្លាំងតាមដែលអ្នកសង្កត់គ្រាប់ចុច\n2. រក្សាចលនាដែលមិនចាំបាច់ជាអប្បបរមា ដើម្បីរក្សាភាពអំណត់របស់អ្នក\n3. ត្រូវចងចាំក្នុងសង្កត់ សញ្ញាភ្លេង ខណៈពេលទុកដៃឱ្យទំនេរ\n4. បង្កើនល្បឿនតាមរយៈរបស់ខាងក្រោម​ដែលអ្នកប្រើ​ ដើម្បីសម្រេចបានកម្រិតសំឡេងបន្ថែម\n\nឆ្លើយដោយលេខនៃជម្រើសត្រឹមត្រូវ (1-4) នៅក្នុង \\boxed{}.",
164
+ "primer_answer": "\boxed{1}"
165
+ },
166
+ "lao_Laoo": {
167
+ "lang_name_native": "ພາສາລາວ",
168
+ "question_label": "ຄຳຖາມ",
169
+ "answer_instruction": "ຕອບດ້ວຍເລກຂອງຕົວເລືອກທີ່ຖືກຕ້ອງ (1-4) ຢູ່ໃນ \\boxed{}.",
170
+ "reason_instruction": "ຄິດ ແລະ ໃຫ້ເຫດຜົນເທື່ອລະຂັ້ນຕອນເປັນພາສາລາວ.",
171
+ "example_reasoning": "ຄຳຖາມຖາມວ່າຄຳແນະນຳໃດບໍ່ຖືກຕ້ອງ. ຂໍ້ຄວາມບອກໃຫ້ຮັກສາມືໃຫ້ສະບາຍແລະຫຼີກລ້ຽງການເຄື່ອນນິ້ວເກີນ, ສະນັ້ນຕົວເລືອກ 2 ແລະ 3 ແມ່ນຄຳແນະນຳທີ່ຖືກຕ້ອງ. ມັນບອກວ່າໃນອາຄໍດີອອນ ສຽງດັງເພີ່ມມາຈາກການໃຊ້ຖົງລົມດ້ວຍຄວາມກົດດັນຫຼືຄວາມໄວຫຼາຍຂຶ້ນ, ສະນັ້ນຕົວເລືອກ 4 ກໍຖືກຕ້ອງເຊັ່ນກັນ. ມັນບອກຢ່າງຈະແຈ້ງວ່າບໍ່ຈຳເປັນຕ້ອງກົດປຸ່ມດ້ວຍແຮງຫຼາຍເພື່ອໃຫ້ສຽງດັງເພີ່ມ, ສະນັ້ນຕົວເລືອກ 1 ແມ່ນຄຳແນະ��ຳທີ່ບໍ່ຖືກຕ້ອງ. ຄຳຕອບແມ່ນ 1.",
172
+ "source": "beam-translated+manual-fix",
173
+ "lid_example": {
174
+ "n_seg": 4,
175
+ "target_share": 1.0,
176
+ "eng_share": 0.0,
177
+ "other_share": 0.0,
178
+ "first_eng_pos": null,
179
+ "top_other": null
180
+ },
181
+ "primer_user": "ຕ້ອງໃຫ້ແນ່ໃຈວ່າມືຂອງທ່ານຜ່ອນຄາຍທີ່ສຸດເທົ່າທີ່ເປັນໄປໄດ້ໃນຂະນະທີ່ທ່ານຍັງກົດປຸ່ມບັນທຶກທັງໝົດຢ່າງຖືກຕ້ອງ ພ້ອມທັງພະຍາຍາມບໍ່ໃຫ້ນິ້ວມືຂອງທ່ານເຄື່ອນໄຫວຫຼາຍເກີນຂອບເຂດ. ວິທີນີ້, ທ່ານຈະເມື່ອຍກັບໂຕເອງໜ້ອຍທີ່ສຸດທີ່ຈະເປັນໄປໄດ້. ຈື່ໄວ້ວ່າບໍ່ຈຳເປັນຕ້ອງກົດປຸ່ມຕ່າງໆດ້ວຍແຮງຄ້າຍຄືກັບເປຍໂນ. ຕາມເງື່ອນໄຂແລ້ວ, ເພື່ອໃຫ້ໄດ້ປະລິມານຫຼາຍຂຶ້ນ ທ່ານຕ້ອງໃຊ້ບັ້ງສູບດ້ວຍແຮງດັນ ຫຼື ຄວາມໄວທີ່ຫຼາຍຂຶ້ນ.\n\nຄຳຖາມ: ອີງໃສ່ບົດອ່ານ, ສິ່ງໃດທີ່ບໍ່ຖືວ່າເປັນຄຳແນະນຳທີ່ຖືກຕ້ອງໃນການຫຼິ້ນແອັກຄໍດ້ຽນທີ່ສຳເລັດ?\n1. ເພື່ອເພີ່ມສຽງ, ເພີ່ມແຮງກົດຢ່າງເວລາທີ່ທ່ານກົດປຸມ\n2. ຮັກສາການເຄື່ອນໄຫວທີ່ບໍ່ຈຳເປັນໃຫ້ນ້ອຍທີ່ສຸດເພື່ອຮັກສາຄວາມແຂງແຮງຂອງທ່ານ\n3. ລະມັດລະວັງເມື່ອກົດສະໝຸດບັນທຶກໃນຂະນະທີ່ຮັກສາຄວາມຜ່ອນຄາຍຢູ່ມື\n4. ເພີ່ມຄວາມໄວໄດ້ເຊິ່ງທ່ານສາມາດດຳເນີນການປ້ຳເຄື່ອງສູບລົມເພື່ອໃຫ້ມີສຽງເພີ່ມຂຶ້ນ\n\nຕອບດ້ວຍເລກຂອງຕົວເລືອກທີ່ຖືກຕ້ອງ (1-4) ຢູ່ໃນ \\boxed{}.",
182
+ "primer_answer": "\boxed{1}"
183
+ },
184
+ "mya_Mymr": {
185
+ "lang_name_native": "မြန်မာ",
186
+ "question_label": "မေးခွန်း",
187
+ "answer_instruction": "မှန်ကန်သော ရွေးချယ်စရာ (1-4) ၏ နံပါတ်ကို \\boxed{} အတွင်း ဖြေပါ။",
188
+ "reason_instruction": "မြန်မာဘာသာဖြင့် တစ်ဆင့်ချင်း စဉ်းစားပြီး အကြောင်းပြချက်ပေးပါ။",
189
+ "example_reasoning": "မေးခွန်းက မည်သည့် အကြံပြုချက်က မမှန်ကန်သလဲ ဟု မေးထားသည်။ စာပိုဒ်က လက်ကို ပေါ့ပေါ့ပါးပါး ထားရန်နှင့် လက်ချောင်း အပိုလှုပ်ရှားမှုကို ရှောင်ရန် ပြောထားသဖြင့် ရွေးချယ်စရာ 2 နှင့် 3 သည် မှန်ကန်သော အကြံပြုချက်များ ဖြစ်သည်။ အကော်ဒီယံတွင် အသံပမာဏ ပိုများလာခြင်းသည် ဘယ်လို့စ်ကို ဖိအားပိုပေးခြင်း သို့မဟုတ် အရှိန်ပိုမြှင့်ခြင်းဖြင့် ရရှိသည်ဟု ပြောထားသဖြင့် ရွေးချယ်စရာ 4 လည်း မှန်ကန်သည်။ အသံပမာဏ ပိုရရန် ကီးများကို အင်အားများများဖြင့် နှိပ်ရန် မလိုဟု တိတိကျကျ ပြောထားသဖြင့် ရွေးချယ်စရာ 1 သည် မမှန်ကန်သော အကြံပြုချက် ဖြစ်သည်။ အဖြေမှာ 1 ဖြစ်သည်။",
190
+ "source": "beam-translated",
191
+ "lid_example": {
192
+ "n_seg": 1,
193
+ "target_share": 1.0,
194
+ "eng_share": 0.0,
195
+ "other_share": 0.0,
196
+ "first_eng_pos": null,
197
+ "top_other": null
198
+ },
199
+ "primer_user": "ခလုတ်အားလုံးကို မှန်ကန်စွာတီးခတ်နေစဉ်တွင် သင်၏လက်ကို တတ်နိုင်သမျှ ဖြေလျှော့ထားပါ - ထို့အတူ လက်ချောင်းများကို မလိုအပ်ဘဲအလွန်အမင်းလှုပ်ရှားခြင်းမျိုးမလုပ်မိရန် ကြိုးစားပါ။ ထိုသို့သောနည်းလမ်းဖြင့်၊ သင်သည် ဖြည်းဖြည်းချင်း ပင်ပန်းလာပါလိမ့်မည်။ ဉပမာပြောရလျှင် စန္ဒယားတီးသည့်အခါ အသံကျယ်ကျယ်ထွက်ရှိရန်အတွက် ခလုတ်ကိုအားများများသုံးကာ ဖိတီးရန် မလိုသည်ကို သိရှိပါ။ အကော်ဒီယံတွင် အသံပိုရရန် အိတ်ကို ဖိအားသို့မဟုတ် အမြန်အသုံးပြုရသည်။\n\nမေးခွန်း: စာပိုဒ်အရ အကော်ဒီယံကိုအောင်မြင်စွာတီးခတ်ခြင်းအတွက် တိကျသော အကြံပြုချက်ဟု မယူဆနိုင်​သော အရာသည် အဘယ်နည်း။\n1. အသံကျယ်ကျယ်ထွက်ရှိရန်အတွက် ခလုတ်များကို အားပိုထည့်၍တီးခြင်း\n2. သင်၏သက်လုံကို ဆက်ထိန်းထားရန် မလိုအပ်ဘဲလှုပ်ရှားခြင်းမျိုးမလုပ်မိရန် ကြိုးစားခြင်း\n3. လက်ကိုဖြေလျှော့ထားလျက် ခလုတ်များတီးခတ်ခြင်းကို သတိထားခြင်း\n4. အသံပိုထွက်ရန် အိတ်ကို ပို၍မြန်မြန်အသုံးပြုခြင်း\n\nမှန်ကန်သော ရွေးချယ်စရာ (1-4) ၏ နံပါတ်ကို \\boxed{} အတွင်း ဖြေပါ။",
200
+ "primer_answer": "\boxed{1}"
201
+ },
202
+ "hin_Deva": {
203
+ "lang_name_native": "हिन्दी",
204
+ "question_label": "प्रश्न",
205
+ "answer_instruction": "सही विकल्प का नंबर (1-4) \\boxed{} के अंदर दें।",
206
+ "reason_instruction": "हिन्दी में चरण-दर-चरण सोचें और तर्क दें।",
207
+ "example_reasoning": "प्रश्न पूछता है कि कौन-सा सुझाव सटीक नहीं है। अनुच्छेद कहता है कि हाथ को आराम से रखें और अतिरिक्त उंगली गति से बचें, इसलिए विकल्प 2 और 3 सटीक सुझाव हैं। यह कहता है कि एकॉर्डियन पर अतिरिक्त आवाज़ बेलोज़ का अधिक दबाव या गति से उपयोग करने से आती है, इसलिए विकल्प 4 भी सटीक है। यह स्पष्ट रूप से कहता है कि अतिरिक्त आवाज़ के लिए कुंजियों को बहुत ज़ोर से दबाने की ज़रूरत नहीं है, इसलिए विकल्प 1 गलत सुझाव है। उत्तर 1 है।",
208
+ "source": "beam-translated",
209
+ "lid_example": {
210
+ "n_seg": 1,
211
+ "target_share": 1.0,
212
+ "eng_share": 0.0,
213
+ "other_share": 0.0,
214
+ "first_eng_pos": null,
215
+ "top_other": null
216
+ },
217
+ "primer_user": "सुनिश्चित करें कि आपका हाथ सभी नोट्स को सही से बजाते हुए भी जितना हो सके, उतना स्थिर है - यह भी प्रयास करें कि आपकी उंगलियाँं अत्यधिक गतिवान न हो। इस तरह, आप जितना हो सकेगा, अपने को थका लेंगे. याद रखें कि ज़्यादा आवाज़ के लिए उसकी कुंजियों को ज़्यादा ज़ोर से दबाने की ज़रूरत नहीं होती है, जैसा कि पियानो में होता है. अकॉर्डियन पर, अतिरिक्त वॉल्यूम प्राप्त करने के लिए, आप कुहनी का उपयोग अधिक दबाव या गति के साथ करते हैं।\n\nप्रश्न: गद्यांश के अनुसार, संगीत के लिए अक्कादियन को सटीक रूप से पुनर्स्थापित क���ने के लिए वास्तव में क्या संभव नहीं है?\n1. अधिक मात्रा में उत्पन्न करने के लिए कुंजियों को अधिक बल के साथ दबाया जाना चाहिए\n2. अपनी उंगलियों से अनावश्यक हरकत करने से बचें ताकि आपकी शक्ति कम न हो\n3. यह सुनिश्चित करने के लिए ध्यान रखा जाना चाहिए कि आपके हाथ अक्काडियन को सही ढंग से पुनः प्राप्त करने के लिए सहज हों\n4. उच्च मात्रा प्राप्त करने के लिए उच्च दबाव या वेग वाले अकॉर्डियन का उपयोग किया जा सकता है\n\nसही विकल्प का नंबर (1-4) \\boxed{} के अंदर दें।",
218
+ "primer_answer": "\boxed{1}"
219
+ },
220
+ "tam_Taml": {
221
+ "lang_name_native": "தமிழ்",
222
+ "question_label": "கேள்வி",
223
+ "answer_instruction": "சரியான விருப்பத்தின் எண்ணை (1-4) \\boxed{} உள்ளே கொடுக்கவும்.",
224
+ "reason_instruction": "தமிழில் படிப்படியாக சிந்தித்து காரணம் கூறவும்.",
225
+ "example_reasoning": "இந்தக் கேள்வி எந்த குறிப்பு துல்லியமானது அல்ல என்பதைக் கேட்கிறது. பத்தியில் கையை தளர்வாக வைத்திருக்கவும், கூடுதல் விரல் அசைவைத் தவிர்க்கவும் என்று கூறப்பட்டுள்ளது; எனவே விருப்பங்கள் 2 மற்றும் 3 துல்லியமான குறிப்புகள். அக்கார்டியனில் கூடுதல் ஒலியளவு பெல்லோஸை அதிக அழுத்தத்துடன் அல்லது வேகத்துடன் பயன்படுத்துவதால் வருகிறது என்று அது கூறுகிறது; எனவே விருப்பம் 4-ம் துல்லியமானது. கூடுதல் ஒலியளவுக்காக விசைகளை அதிக விசையுடன் அடிக்க வேண்டிய அவசியம் இல்லை என்று அது வெளிப்படையாகக் கூறுகிறது; எனவே விருப்பம் 1 துல்லியமற்ற குறிப்பு. பதில் 1.",
226
+ "source": "beam-translated",
227
+ "lid_example": {
228
+ "n_seg": 4,
229
+ "target_share": 1.0,
230
+ "eng_share": 0.0,
231
+ "other_share": 0.0,
232
+ "first_eng_pos": null,
233
+ "top_other": null
234
+ },
235
+ "primer_user": "அனைத்து சுரங்களையும் சரியாக வாசிக்கும் போது உன் கை முடிந்தவரை சௌகரியமாக இருப்பதை உறுதி செய்து கொள்ளவும், மேலும், விரல்களால் அதிக தேவையற்ற அசைவுகளை செய்யாமல் இருக்க முயலவும். இந்த வழியில், முடிந்த அளவு குறைவாக சோர்வடைவீர்கள். அதிக ஒலிக்காக விசைகளை பியானோவில் அழுத்துவது போல அதிக ஆற்றலுடன் அழுத்தத் தேவையில்லை என்பதை நினைவில் கொள்ளுங்கள். அக்கார்டியனில் அதிக ஒலியினைப் பெறுவதற்கு, நீங்கள் அதிக அழுத்தம் அல்லது வேகமுடைய துருத்திகளைப் பயன்படுத்தலாம்.\n\nகேள்வி: கொடுக்கப்பட்டிருக்கும் இந்த பத்தியில், சுரங்களை துல்லியமாக வாசிப்பதற்கான சரியான உதவிக் குறிப்பு இல்லை எனக் கருதப்படுவது எது?\n1. அதிக அளவு ஒலியினை ஏற்படுத்த, விசைகளை அதிக அளவு ஆற்றலுடன் அழுத்த வேண்டும்\n2. உங்கள் ஸ்டெமினா குறையாமல் இருக்க விரல்களால் தேவையற்ற அசைவுகளைச் செய்யாமல் இருக்க வேண்டும்\n3. சுரங்களை சரியாக வாசிக்கும் போது உங்கள் கைகளை சௌகரியமாக வைத்திருக்கிறீர்களா என்பதில் கவனம் செலுத்த வேண்டும்\n4. அதிக அளவு ஒலியினைப் பெறுவதற்கு, அதிக அழுத்தத்தையோ அல்லது வேகத்தையோ கொண்ட துருத்திகளைப் பயன்படுத்தலாம்\n\nசரியான விருப்பத்தின் எண்ணை (1-4) \\boxed{} உள்ளே கொடுக்கவும்.",
236
+ "primer_answer": "\boxed{1}"
237
+ },
238
+ "kor_Hang": {
239
+ "lang_name_native": "한국어",
240
+ "question_label": "질문",
241
+ "answer_instruction": "정답 선택지의 번호(1-4)를 \\boxed{} 안에 입력하세요.",
242
+ "reason_instruction": "한국어로 단계별로 생각하고 추론하세요.",
243
+ "example_reasoning": "이 문제는 어떤 팁이 정확하지 않은지 묻습니다. 지문은 손을 편안하게 유지하고 불필요한 손가락 움직임을 피하라고 하므로, 선택지 2와 3은 정확한 팁입니다. 또한 아코디언에서는 벨로우즈를 더 강하게 누르거나 빠르게 사용해 음량을 높인다고 하므로, 선택지 4도 정확합니다. 지문은 더 큰 음량을 위해 건반을 세게 칠 필요가 없다고 명시하므로, 선택지 1이 부정확한 팁입니다. 정답은 1입니다.",
244
+ "source": "beam-translated",
245
+ "lid_example": {
246
+ "n_seg": 4,
247
+ "target_share": 1.0,
248
+ "eng_share": 0.0,
249
+ "other_share": 0.0,
250
+ "first_eng_pos": null,
251
+ "top_other": null
252
+ },
253
+ "primer_user": "모든 음을 올바르게 누른 상태에서 손을 최대한 편하게 둡니다. 또한 손가락으로 관련 없는 동작을 많이 하지 않도록 하십시오. 이 방식을 통해 연주의 피로도를 최대한 낮출 수 있습니다. 피아노와 달리, 아코디언은 더 큰 소리를 내기 위해 건반을 세게 칠 필요가 없다는 것을 기억하세요. 아코디언에 볼륨을 더 주려면 풀무에 힘을 더 주거나 빠르게 풀무를 사용해야 합니다.\n\n질문: 지문에 따르면, 아코디언을 잘 연주하기 위한 정확한 팁으로 볼 수 없는 것은 무엇입니까?\n1. 볼륨을 높이려면 키를 더 세게 누른다\n2. 체력을 유지하기 위해 불필요한 움직임을 최소화한다\n3. 손을 최대한 편하게 유지하면서 정확한 음표를 치는 데 신경을 쓴다\n4. 볼륨을 추가로 높이려면 풀무를 작동하는 속도를 높인다\n\n정답 선택지의 번호(1-4)를 \\boxed{} 안에 입력하세요.",
254
+ "primer_answer": "\boxed{1}"
255
+ },
256
+ "jpn_Jpan": {
257
+ "lang_name_native": "日本語",
258
+ "question_label": "質問",
259
+ "answer_instruction": "正解の選択肢の番号(1-4)を\\boxed{}の中に入れて答えてください。",
260
+ "reason_instruction": "日本語で段階的に考えて推論してください。",
261
+ "example_reasoning": "この質問は、どのヒントが正確ではないかを問うています。文章では、手をリラックスさせ、余分な指の動きを避けるように述べているため、選択肢2と3は正確なヒントです。また、アコーディオンでは余分な音量はベローズをより強くまたは速く使うことから生まれると述べているため、選択肢4も正確です。さらに、余分な音量のために鍵盤を強く叩く必要はないと明確に述べているため、選択肢1が不正確なヒントです。答えは1です。",
262
+ "source": "beam-translated",
263
+ "lid_example": {
264
+ "n_seg": 1,
265
+ "target_share": 1.0,
266
+ "eng_share": 0.0,
267
+ "other_share": 0.0,
268
+ "first_eng_pos": null,
269
+ "top_other": null
270
+ },
271
+ "primer_user": "音符どおりに正確に奏でながら、手をできるだけリラックスした状態で保つようにしましょう。また、指が余計な動きをしないように努めましょう。 そうすることで、疲れがたまりにくくなります。ピアノのように音量を出すために力を入れて鍵盤を叩く必要はないことを覚えておいてください。 アコーディオンを弾くときに音量を上げるには、蛇腹を強く、または速く動かします。\n\n質問: 前述の内容によると、上手にアコーディオンを弾くための適切なコツとして考えられないものはどれでしょう?\n1. 音量を上げるために、鍵盤を叩く力を強くする\n2. 体力を温存するために、無駄な動きを最小限に抑える\n3. 手をリラックスした状態に保ちながら、音���通りに弾くように注意する\n4. 音量を上げるために蛇腹を動かす速度を上げる\n\n正解の選択肢の番号(1-4)を\\boxed{}の中に入れて答えてください。",
272
+ "primer_answer": "\boxed{1}"
273
+ },
274
+ "arb_Arab": {
275
+ "lang_name_native": "العربية",
276
+ "question_label": "سؤال",
277
+ "answer_instruction": "أجب برقم الخيار الصحيح (1-4) داخل \\boxed{}.",
278
+ "reason_instruction": "فكر وعلل خطوة بخطوة بالعربية.",
279
+ "example_reasoning": "يسأل السؤال عن أي نصيحة غير دقيقة. يقول النص إن należy إبقاء اليد مسترخية وتجنب الحركة الإضافية للأصابع، لذا فإن الخيارات 2 و3 نصائح دقيقة. ويقول إن في الأكورديون يأتي الحجم الإضافي من استخدام fuelle مع ضغط أو سرعة أكبر، لذا فإن الخيار 4 دقيق أيضاً. ويقول صراحةً إن لا حاجة لضرب المفاتيح بقوة كبيرة للحصول على حجم إضافي، لذا فإن الخيار 1 هو النصيحة غير الدقيقة. الإجابة هي 1.",
280
+ "source": "beam-translated",
281
+ "lid_example": {
282
+ "n_seg": 4,
283
+ "target_share": 1.0,
284
+ "eng_share": 0.0,
285
+ "other_share": 0.0,
286
+ "first_eng_pos": null,
287
+ "top_other": null
288
+ },
289
+ "primer_user": "تأكد من استرخاء يدك قدر الإمكان مع الاستمرار في ضرب كل النغمات بشكل صحيح - حاول كذلك عدم القيام بحركاتٍ غريبةٍ بأصابعك. لن تبذل مجهوداً كبيراً إذا اتبعت تلك الطريقة. ضع نصب عينيك أنه ليس عليك الضغط على مفاتيح الأكورديون بقوةٍ كبيرةٍ لتحصل على مستوى صوتٍ أعلى كما في البيانو. في الأكورديون، للحصول على صوت أعلى إضافي، يمكنك استخدام المنفاخ بمزيد من الضغط أو السرعة.\n\nسؤال: وفقاً للفقرة، ما الذي لا يُعتبر نصيحة دقيقة للعزف الناجح على الأكورديون؟\n1. للحصول على صوت أعلى، زدّ القوة التي تضغط بها على المفاتيح\n2. أبقِ الحركة غير الضرورية في حدها الأدنى للحفاظ على طاقتك\n3. احرص على ضرب النغمات مع إبقاء اليد مسترخية\n4. زدّ السرعة التي تُشغل بها المنفاخ للحصول على صوت إضافي\n\nأجب برقم الخيار الصحيح (1-4) داخل \\boxed{}.",
290
+ "primer_answer": "\boxed{1}"
291
+ },
292
+ "yor_Latn": {
293
+ "lang_name_native": "Yorùbá",
294
+ "question_label": "Ìbéèrè",
295
+ "answer_instruction": "Dáhùn pẹ̀lú nọ́mbà àṣàyàn tó tọ́ (1-4) nínú \\boxed{}.",
296
+ "reason_instruction": "Ronú kí o sì ṣe àlàyé lẹ́kún rẹ̀ ní Yorùbá.",
297
+ "example_reasoning": "Ìbéèrè náà béèrè àbá wo ni kò tọ́. Àpilẹ̀kọ náà sọ pé kí ọwọ́ wà ní ìsinmi kí a sì yẹra fún ìrìn ìka afikún, nítorí náà àbá 2 àti 3 jẹ́ àbá tó tọ́. Ó sọ pé lórí àkọ́dìọ̀nì, ìwọn ohùn afikún wá láti lílo bellows pẹ̀lú ìfúnpá tàbí ìyára jù, nítorí náà àbá 4 náà tọ́. Ó sọ kedere pé kò sí ìdí láti lu àwọn kọ́kọ́rọ́ pẹ̀lú agbára púpọ̀ fún ìwọn ohùn afikún, nítorí náà àbá 1 ni àbá tó kò tọ́. Ìdáhùn náà ni 1.",
298
+ "source": "beam-translated",
299
+ "lid_example": {
300
+ "n_seg": 4,
301
+ "target_share": 1.0,
302
+ "eng_share": 0.0,
303
+ "other_share": 0.0,
304
+ "first_eng_pos": null,
305
+ "top_other": null
306
+ },
307
+ "primer_user": "Rí i dájú pé ọwọ́ rẹ ní ìsinmi bó se yẹ nígbàtí o bá ń kọlu gbogbo àkíyèsí tí ó tó - pẹ̀lú pé ki o má dààmú ọmọ ìka ọwọ́ rẹ jùlọ. Bó ba n ṣe báyí kò ní rẹ̀ ẹ́ jù. Rántí pé o kò níláti fi agbára tẹ ojú rẹ̀ kí ohun rẹ̀ ga síi bí ti piano. Láti mú ohun akodioni tóbi si wàá lò àwọn belò pẹ̀lú agbára tàbí ere sí i.\n\nÌbéèrè: Gẹgẹ bi oju ewe naa, kini a o pe ni ojulowo itọni fun imọ titẹ àkọdiọnu daadaa??\n1. Fun afikun iwọn didun, lo agbara pipọ pẹlu eyiti o lu awọn bọtini naa\n2. Maṣe maa rin ainidi kiri ki o le ni okun si\n3. Ṣe akiyesi lilu awọn nooti lakoko mimu ọwọ rẹ rọ\n4. Mu iyara pọ si pẹlu eyi ti o fi ṣẹ awọn ohun agbara isalẹ lati ṣaṣeyọri iwọn didun afikun\n\nDáhùn pẹ̀lú nọ́mbà àṣàyàn tó tọ́ (1-4) nínú \\boxed{}.",
308
+ "primer_answer": "\boxed{1}"
309
+ },
310
+ "plt_Latn": {
311
+ "lang_name_native": "fiteny malagasy",
312
+ "question_label": "Fanontaniana",
313
+ "answer_instruction": "Valio amin'ny laharan'ny safidy marina (1-4) ao anatin'ny \\boxed{}.",
314
+ "reason_instruction": "Saina sy araraotra tsikelikely amin'ny fiteny malagasy.",
315
+ "example_reasoning": "Ny fanontaniana dia manontany izay toro-hevitra TSY marina. Milaza ny lahatsoratra fa tokony hampitony ny tanana ary hisoroka ny hetsika fanampiny amin'ny rantsan-tanana, ka marina ny safidy 2 sy 3. Milaza izy fa amin'ny accordion dia avy amin'ny fampiasana ny bellows amin'ny tsindry na hafainganana bebe kokoa ny feo fanampiny, ka marina koa ny safidy 4. Milaza mazava izy fa tsy ilaina ny manapoka ny kitso amin'ny herim-po be ho an'ny feo fanampiny, ka ny safidy 1 no toro-hevitra tsy marina. Ny valiny dia 1.",
316
+ "source": "beam-translated",
317
+ "lid_example": {
318
+ "n_seg": 4,
319
+ "target_share": 1.0,
320
+ "eng_share": 0.0,
321
+ "other_share": 0.0,
322
+ "first_eng_pos": null,
323
+ "top_other": null
324
+ },
325
+ "primer_user": "Ataovy azo antoka fa milamina tsara ny tanana eo am-panoratana ny naoty araka ny tokony ho izy - ary koa miezaha tsy dia manao fihetsika be amin’ny ratsan-tananao. Amin’ity fomba ity dia mba tsy ho reraka be ny tenanao. Tadidio fa tsy ilaina ny manindry ny bokotra mafy mba hahazoana feo mafy toy ny amin’ny piano. Amin’ny angorodao, ahazoana feo mafy tsara, dia tsindrina mafimafy ary haingana kokoa ny lakolosy.\n\nFanontaniana: Araka ilay sombin-dahatsoratra, inona no tsy heverina ho torohevitra mitombina hitendrena angorodao am-pahombiazana?\n1. Hampiakarana ny feony, ampitomboy ny hery amin’ny fanindrinao ny bokotra\n2. Ahenao ny fihetsika tsy ilaina mba hitazomana ny tanjakao\n3. Mifantoha lalina amin’ny naoty sady hazony mba ho malefaka hatrany ny tanana\n4. Ampitomboy ny hafainganana amin’ny tafoforana mba ahatrarana ny feo avo indrindra\n\nValio amin'ny laharan'ny safidy marina (1-4) ao anatin'ny \\boxed{}.",
326
+ "primer_answer": "\boxed{1}"
327
+ }
328
+ }
code/relid.py ADDED
@@ -0,0 +1,48 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Robustness pass (no inference): for every atlas/math/rewrite row add
2
+ lid_noquote GlotLID profile after dropping segments that quote the prompt (passage/options/problem)
3
+ lingua_eng English share by an independent detector (Lingua)
4
+ lingua_eng_noquote the same, quote-stripped
5
+ Rows are rewritten in place. Usage: python relid.py"""
6
+ import json, os
7
+ from concurrent.futures import ThreadPoolExecutor
8
+ from datasets import load_dataset
9
+ from lid import profile, lingua_eng_share
10
+ from atlas import LANGS as ATLAS_LANGS, mc_prompt
11
+
12
+ tps = json.load(open("data/atlas/templates.json", encoding="utf-8"))
13
+ par = json.load(open("data/math/parallel.json", encoding="utf-8"))
14
+ bel = {}
15
+
16
+
17
+ def belebele(lang):
18
+ if lang not in bel:
19
+ bel[lang] = {(r["link"], r["question_number"]): r for r in load_dataset("facebook/belebele", lang, split="test")}
20
+ return bel[lang]
21
+
22
+
23
+ def prompt_of(r):
24
+ if "link" in r:
25
+ return mc_prompt(belebele(r["lang"])[(r["link"], r["qnum"])], tps[r["lang"]])
26
+ return par[str(r["id"])][r["lang"]]["text"]
27
+
28
+
29
+ def enrich(r):
30
+ p = prompt_of(r)
31
+ r["lid_noquote"] = profile(r["reasoning"], r["lang"], exclude=p)
32
+ r["lingua_eng"] = lingua_eng_share(r["reasoning"])
33
+ r["lingua_eng_noquote"] = lingua_eng_share(r["reasoning"], exclude=p)
34
+ return r
35
+
36
+
37
+ for path in ("data/atlas/runs_low.jsonl", "data/atlas/runs_high.jsonl", "data/math/runs_low.jsonl",
38
+ "data/math/runs_high.jsonl"):
39
+ if not os.path.exists(path):
40
+ continue
41
+ rows = [json.loads(l) for l in open(path, encoding="utf-8")]
42
+ for lang in {r["lang"] for r in rows if "link" in r}:
43
+ belebele(lang) # load once, single-threaded
44
+ with ThreadPoolExecutor(4) as ex:
45
+ rows = list(ex.map(enrich, rows))
46
+ with open(path, "w", encoding="utf-8") as f:
47
+ f.writelines(json.dumps(r, ensure_ascii=False) + "\n" for r in rows)
48
+ print("enriched", path, len(rows))
code/thinking.py ADDED
@@ -0,0 +1,60 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Babak 3: build in-language reasoning traces where Beam got the answer right but thought in English.
2
+ Beam rewrites its own correct reasoning into the target language (as visible content, which follows language
3
+ instructions far better than reasoning does). Kept only if GlotLID says >= 0.8 target language AND every number
4
+ of the original reasoning survives. Output: data/thinking/rewritten.jsonl (trace_source='rewritten').
5
+ Usage: python thinking.py --per-lang 60"""
6
+ import argparse, json, os, re
7
+ from beam import chat, run_many, spent
8
+ from lid import profile
9
+
10
+ LANGS = ["ind_Latn", "jav_Latn", "sun_Latn", "zsm_Latn", "tgl_Latn", "ceb_Latn", "vie_Latn", "tha_Thai"]
11
+ # v2: v1 let Beam *solve* the embedded problem (251/439 outputs were just \boxed{answer}); now the text is fenced
12
+ # and the task is stated as translation-only.
13
+ PROMPT = """You are a translator. Translate the text inside <notes> into natural, fluent {name}.
14
+ This is a translation task only: do NOT solve anything, do NOT add a final answer, do NOT shorten.
15
+ Keep every step, number and calculation exactly as written. Output only the translated text, without the tags.
16
+
17
+ <notes>
18
+ {reasoning}
19
+ </notes>"""
20
+
21
+
22
+ def nums(t):
23
+ t = re.sub(r"(?<=\d)[.,](?=\d{3}\b)", "", t) # drop thousands separators (1.250.000 / 1,250,000)
24
+ return set(re.findall(r"\d+", t)) # ponytail: digit-run set; tolerant to 1,250 vs 1250 formatting
25
+
26
+
27
+ def main():
28
+ ap = argparse.ArgumentParser(); ap.add_argument("--per-lang", type=int, default=60); a = ap.parse_args()
29
+ names = {c: v["lang_name_native"] for c, v in json.load(open("data/atlas/templates.json", encoding="utf-8")).items()}
30
+ par = json.load(open("data/math/parallel.json", encoding="utf-8"))
31
+ runs = [json.loads(l) for l in open("data/math/runs_low.jsonl", encoding="utf-8")]
32
+ pool = {}
33
+ for r in runs: # correct, reasoning mostly NOT in the target language
34
+ ts = r["lid"]["target_share"]
35
+ if r["lang"] in LANGS and r["correct"] and r["reasoning"] and (ts is None or ts < 0.8):
36
+ pool.setdefault(r["lang"], {}).setdefault(r["id"], r) # one trace per problem per language
37
+ jobs = [r for l in LANGS for r in list(pool.get(l, {}).values())[:a.per_lang]]
38
+
39
+ def one(r):
40
+ out = chat([{"role": "user", "content": PROMPT.format(name=names[r["lang"]], reasoning=r["reasoning"])}],
41
+ "low", 6000, temperature=0.2, tag="rewrite-v2")
42
+ new = out["content"].strip()
43
+ lid = profile(new, r["lang"])
44
+ a, b = nums(new), nums(r["reasoning"])
45
+ jac = len(a & b) / max(1, len(a | b)) # number-set Jaccard; translation may re-spell a few ordinals
46
+ ok = (lid["target_share"] or 0) >= 0.8 and jac >= 0.9
47
+ return {"lang": r["lang"], "id": r["id"], "answer": r["answer"], "effort": r["effort"],
48
+ "user": par[str(r["id"])][r["lang"]]["text"], "reasoning": new, "content": r["content"],
49
+ "correct": ok, "lid": lid, "numbers_jaccard": round(jac, 3),
50
+ "original_reasoning_lid": r["lid"], "trace_source": "rewritten"}
51
+
52
+ rows = [x for _, x in run_many(one, jobs, 10, "rewrite") if not isinstance(x, Exception)]
53
+ os.makedirs("data/thinking", exist_ok=True)
54
+ with open("data/thinking/rewritten.jsonl", "w", encoding="utf-8") as f:
55
+ f.writelines(json.dumps(x, ensure_ascii=False) + "\n" for x in rows)
56
+ print(f"{sum(x['correct'] for x in rows)}/{len(rows)} rewrites kept; spent ${spent()}")
57
+
58
+
59
+ if __name__ == "__main__":
60
+ main()
data/atlas_runs.jsonl CHANGED
The diff for this file is too large to render. See raw diff
 
data/math_runs.jsonl CHANGED
The diff for this file is too large to render. See raw diff
 
data/prompt_templates.json CHANGED
@@ -5,7 +5,9 @@
5
  "answer_instruction": "Answer with the number of the correct option (1-4) inside \\boxed{}.",
6
  "reason_instruction": "Think and reason step by step in English.",
7
  "example_reasoning": "The question asks which tip is NOT accurate. The passage says to keep the hand relaxed and to avoid extra finger motion, so options 2 and 3 are accurate tips. It says that on the accordion extra volume comes from using the bellows with more pressure or speed, so option 4 is also accurate. It explicitly says there is no need to hit the keys with a lot of force for extra volume, so option 1 is the inaccurate tip. The answer is 1.",
8
- "source": "human"
 
 
9
  },
10
  "ind_Latn": {
11
  "lang_name_native": "bahasa Indonesia",
@@ -13,7 +15,9 @@
13
  "answer_instruction": "Jawab dengan nomor pilihan yang benar (1-4) di dalam \\boxed{}.",
14
  "reason_instruction": "Berpikir dan bernalarlah langkah demi langkah dalam bahasa Indonesia.",
15
  "example_reasoning": "Pertanyaannya menanyakan tips mana yang TIDAK tepat. Bacaan mengatakan tangan harus tetap rileks dan jari tidak banyak bergerak, jadi pilihan 2 dan 3 adalah tips yang tepat. Bacaan juga mengatakan bahwa pada akordeon volume tambahan didapat dengan menekan atau menggerakkan bellow lebih kuat atau lebih cepat, jadi pilihan 4 juga tepat. Bacaan dengan jelas mengatakan tidak perlu menekan tuts dengan keras untuk menambah volume, jadi pilihan 1 adalah tips yang tidak tepat. Jawabannya 1.",
16
- "source": "human"
 
 
17
  },
18
  "jav_Latn": {
19
  "lang_name_native": "Basa Jawa",
@@ -29,7 +33,9 @@
29
  "other_share": 0.0,
30
  "first_eng_pos": null,
31
  "top_other": null
32
- }
 
 
33
  },
34
  "sun_Latn": {
35
  "lang_name_native": "Basa Sunda",
@@ -45,7 +51,9 @@
45
  "other_share": 0.0,
46
  "first_eng_pos": null,
47
  "top_other": null
48
- }
 
 
49
  },
50
  "zsm_Latn": {
51
  "lang_name_native": "Bahasa Melayu",
@@ -61,7 +69,9 @@
61
  "other_share": 0.0,
62
  "first_eng_pos": null,
63
  "top_other": null
64
- }
 
 
65
  },
66
  "tgl_Latn": {
67
  "lang_name_native": "Tagalog",
@@ -77,7 +87,9 @@
77
  "other_share": 0.0,
78
  "first_eng_pos": null,
79
  "top_other": null
80
- }
 
 
81
  },
82
  "ceb_Latn": {
83
  "lang_name_native": "Sinugbuanong Binisaya",
@@ -93,7 +105,9 @@
93
  "other_share": 0.0,
94
  "first_eng_pos": null,
95
  "top_other": null
96
- }
 
 
97
  },
98
  "vie_Latn": {
99
  "lang_name_native": "Tiếng Việt",
@@ -109,7 +123,9 @@
109
  "other_share": 0.0,
110
  "first_eng_pos": null,
111
  "top_other": null
112
- }
 
 
113
  },
114
  "tha_Thai": {
115
  "lang_name_native": "ไทย",
@@ -125,7 +141,9 @@
125
  "other_share": 0.0,
126
  "first_eng_pos": null,
127
  "top_other": null
128
- }
 
 
129
  },
130
  "khm_Khmr": {
131
  "lang_name_native": "ខ្មែរ",
@@ -141,7 +159,9 @@
141
  "other_share": 0.0,
142
  "first_eng_pos": null,
143
  "top_other": null
144
- }
 
 
145
  },
146
  "lao_Laoo": {
147
  "lang_name_native": "ພາສາລາວ",
@@ -157,7 +177,9 @@
157
  "other_share": 0.0,
158
  "first_eng_pos": null,
159
  "top_other": null
160
- }
 
 
161
  },
162
  "mya_Mymr": {
163
  "lang_name_native": "မြန်မာ",
@@ -173,7 +195,9 @@
173
  "other_share": 0.0,
174
  "first_eng_pos": null,
175
  "top_other": null
176
- }
 
 
177
  },
178
  "hin_Deva": {
179
  "lang_name_native": "हिन्दी",
@@ -189,7 +213,9 @@
189
  "other_share": 0.0,
190
  "first_eng_pos": null,
191
  "top_other": null
192
- }
 
 
193
  },
194
  "tam_Taml": {
195
  "lang_name_native": "தமிழ்",
@@ -205,7 +231,9 @@
205
  "other_share": 0.0,
206
  "first_eng_pos": null,
207
  "top_other": null
208
- }
 
 
209
  },
210
  "kor_Hang": {
211
  "lang_name_native": "한국어",
@@ -221,7 +249,9 @@
221
  "other_share": 0.0,
222
  "first_eng_pos": null,
223
  "top_other": null
224
- }
 
 
225
  },
226
  "jpn_Jpan": {
227
  "lang_name_native": "日本語",
@@ -237,7 +267,9 @@
237
  "other_share": 0.0,
238
  "first_eng_pos": null,
239
  "top_other": null
240
- }
 
 
241
  },
242
  "arb_Arab": {
243
  "lang_name_native": "العربية",
@@ -253,7 +285,9 @@
253
  "other_share": 0.0,
254
  "first_eng_pos": null,
255
  "top_other": null
256
- }
 
 
257
  },
258
  "yor_Latn": {
259
  "lang_name_native": "Yorùbá",
@@ -269,7 +303,9 @@
269
  "other_share": 0.0,
270
  "first_eng_pos": null,
271
  "top_other": null
272
- }
 
 
273
  },
274
  "plt_Latn": {
275
  "lang_name_native": "fiteny malagasy",
@@ -285,6 +321,8 @@
285
  "other_share": 0.0,
286
  "first_eng_pos": null,
287
  "top_other": null
288
- }
 
 
289
  }
290
  }
 
5
  "answer_instruction": "Answer with the number of the correct option (1-4) inside \\boxed{}.",
6
  "reason_instruction": "Think and reason step by step in English.",
7
  "example_reasoning": "The question asks which tip is NOT accurate. The passage says to keep the hand relaxed and to avoid extra finger motion, so options 2 and 3 are accurate tips. It says that on the accordion extra volume comes from using the bellows with more pressure or speed, so option 4 is also accurate. It explicitly says there is no need to hit the keys with a lot of force for extra volume, so option 1 is the inaccurate tip. The answer is 1.",
8
+ "source": "human",
9
+ "primer_user": "Make sure your hand is as relaxed as possible while still hitting all the notes correctly - also try not to make much extraneous motion with your fingers. This way, you will tire yourself out as little as possible. Remember there's no need to hit the keys with a lot of force for extra volume like on the piano. On the accordion, to get extra volume, you use the bellows with more pressure or speed.\n\nQuestion: According to the passage, what would not be considered an accurate tip for successfully playing the accordion?\n1. For additional volume, increase the force with which you hit the keys\n2. Keep unnecessary movement to a minimum in order to preserve your stamina\n3. Be mindful of hitting the notes while maintaining a relaxed hand\n4. Increase the speed with which you operate the bellows to achieve extra volume\n\nAnswer with the number of the correct option (1-4) inside \\boxed{}.",
10
+ "primer_answer": "\boxed{1}"
11
  },
12
  "ind_Latn": {
13
  "lang_name_native": "bahasa Indonesia",
 
15
  "answer_instruction": "Jawab dengan nomor pilihan yang benar (1-4) di dalam \\boxed{}.",
16
  "reason_instruction": "Berpikir dan bernalarlah langkah demi langkah dalam bahasa Indonesia.",
17
  "example_reasoning": "Pertanyaannya menanyakan tips mana yang TIDAK tepat. Bacaan mengatakan tangan harus tetap rileks dan jari tidak banyak bergerak, jadi pilihan 2 dan 3 adalah tips yang tepat. Bacaan juga mengatakan bahwa pada akordeon volume tambahan didapat dengan menekan atau menggerakkan bellow lebih kuat atau lebih cepat, jadi pilihan 4 juga tepat. Bacaan dengan jelas mengatakan tidak perlu menekan tuts dengan keras untuk menambah volume, jadi pilihan 1 adalah tips yang tidak tepat. Jawabannya 1.",
18
+ "source": "human",
19
+ "primer_user": "Pastikan tangan dalam keadaan serileks mungkin sambil tetap menekan setiap nada dengan benar - Upayakan jari tidak membuat banyak gerakan ekstra. Dengan cara ini, Anda akan mengurangi rasa lelah Anda. Ingatlah bahwa tidak perlu menekan tuts terlalu keras untuk mendapatkan volume ekstra seperti pada piano. Di akordion, untuk mendapatkan volume lebih besar, Anda menggunakan ububan dengan tekanan atau kecepatan lebih besar.\n\nPertanyaan: Menurut kutipan tersebut, apa yang bukan merupakan tips akurat untuk memainkan akordion dengan sukses?\n1. Untuk volume yang lebih keras, tingkatkan kekuatan tekanan yang Anda gunakan untuk menekan tuts\n2. Buat seminimal mungkin gerakan yang tidak diperlukan untuk menjaga stamina Anda\n3. Perhatikan saat menekan nada sambil mempertahankan tangan yang rileks\n4. Tingkatkan kecepatan Anda dalam mengoperasikan ububan untuk suara yang lebih keras\n\nJawab dengan nomor pilihan yang benar (1-4) di dalam \\boxed{}.",
20
+ "primer_answer": "\boxed{1}"
21
  },
22
  "jav_Latn": {
23
  "lang_name_native": "Basa Jawa",
 
33
  "other_share": 0.0,
34
  "first_eng_pos": null,
35
  "top_other": null
36
+ },
37
+ "primer_user": "Mesthekake tanganmu anteng kambi isih nabuh kabeh calathu kanthi bener - uga coba driji-drijimu aja kakehan obah. Kanthi cara iki, sampeyan bakal ora krasa kesel banget. Kudu dieling-eling yen sampeyan ora kudu nggunakake tenaga sing gedhe kanggo mijet kunci kuwi kaya mijet piano. Ing akordian, kanggo entuk volume tambahan, sampeyan nggunakake embusan kanthi luwih mlenet lan cepet.\n\nPitakonan: Miturut pethikan kasebut, apa sing ora bisa dianggep tips ingkang bener kanggo dolan akordian?\n1. Kanggo volume tambahan, nggunakake tenaga sing gedhe kanggo mijet kunci\n2. Aja kakehan obah kanggo njaga stamina\n3. Mesthekake wektu nabuh calathu tangane anteng\n4. Nggunakake embusan kanthi luwih cepet kanggo entuk volume tambahan\n\nWangsulana kanthi nomer pilihan sing bener (1-4) ing njero \\boxed{}.",
38
+ "primer_answer": "\boxed{1}"
39
  },
40
  "sun_Latn": {
41
  "lang_name_native": "Basa Sunda",
 
51
  "other_share": 0.0,
52
  "first_eng_pos": null,
53
  "top_other": null
54
+ },
55
+ "primer_user": "Pastikeun leungeun anjeun sakumaha tiasa santai bari tetep pencét kabéh nada anu leres - ogé cobi ulah ngadamel seueur gerakan asing sareng ramo anjeun. Ku cara ieu, anjeun bakal capék padahal ngan maén sakedik. Émut, henteu kudu mencet konci sakuat tanaga pikeun nambah suara sapertos dina piano. Dinu akordion, pikeun ngengingkeun volume langkung, anggo bagian handap langkung teuas atanapi langkung enggal.\n\nPatarosan: Dumasar kana bagéan tina karangan anu disayogikeun, piwuruk mana anu teu luyu patali cara anu bener maénkeun akordion?\n1. Ngarah nambah tarik sorana, pencét konci anu diteken leuwih teuas\n2. Ulah seueur teuing gerak teu perlu pikeun ngajaga stamina anjeun\n3. Ati-ati dina mencét koncina bari tetep leungeun mah santai\n4. Paénkeun bagian lipetan akordion leuwih gancang ngarah volume langkung tarik deui\n\nJawab ku nomer pilihan anu bener (1-4) di jero \\boxed{}.",
56
+ "primer_answer": "\boxed{1}"
57
  },
58
  "zsm_Latn": {
59
  "lang_name_native": "Bahasa Melayu",
 
69
  "other_share": 0.0,
70
  "first_eng_pos": null,
71
  "top_other": null
72
+ },
73
+ "primer_user": "Pastikan tangan anda setenang yang boleh sementara masih menekan semua tanda dengan betul - selain itu, cuba untuk tidak membuat gerakan luaran dengan jari anda. Dengan cara ini, anda akan memenatkan diri sendiri sesedikit yang mungkin. Ingatlah bahawa tidak perlu menggunakan banyak tenaga untuk menekan kekunci untuk mendapat bunyi yang kuat seperti piano. Pada akordion, bagi mendapatkan kelantangan tambahan, anda guna pakai belos pada tekanan atau kelajuan yang lebih tinggi.\n\nSoalan: Menurut petikan, apakah petua yang dianggap tidak tepat untuk memainkan akordion dengan baik?\n1. Untuk menambah kelantangan, tingkatkan daya anda menekan kekunci\n2. Minimumkan pergerakan yang tidak diperlukan untuk menyimpan stamina anda\n3. Pastikan anda mencapai not sambil memastikan jari tenang\n4. Tingkatkan kelajuan anda mengendalikan belos untuk mendapatkan lebih kelantangan\n\nJawab dengan nombor pilihan yang betul (1-4) di dalam \\boxed{}.",
74
+ "primer_answer": "\boxed{1}"
75
  },
76
  "tgl_Latn": {
77
  "lang_name_native": "Tagalog",
 
87
  "other_share": 0.0,
88
  "first_eng_pos": null,
89
  "top_other": null
90
+ },
91
+ "primer_user": "Siguruhing kalmado ang iyong kamay hangga't posible habang wastong inaabot pa rin ang lahat ng mga nota - subukan ding huwag gumawa ng labis na paggalaw sa iyong mga daliri. Sa ganitong paraan, papagurin mo lamang ang sarili mo nang kaunti hangga't maaari. Tandaan, walang pangangailangan na pindutin nang madiin ang mga teklado para lumakas ang tunog katulad ng kapag sa piyano. Sa akordyon, upang makakuha ng karagdagang dami, gumamit ka ng mga bellow na mayroong higit na presyon at bilis.\n\nTanong: Ayon sa sipi, ano ang hindi maituturing na tumpak na paalala para sa matagumpay na pagpapatugtog ng akordyon?\n1. Para sa mas malakas na volume, dagdagan ang puwersa ng pagpindot sa teklado\n2. Hanggang maaari, iwasan ang hindi kinakailangang paggalaw upang mapreserba ang iyong stamina\n3. Maging maingat sa pag-abot ng nota habang nananatiling kalmado ang kamay\n4. Mas bilisan pa ang paggalaw ng mga bellow upang mas mapalakas ang volume\n\nSagutin gamit ang numero ng tamang opsyon (1-4) sa loob ng \\boxed{}.",
92
+ "primer_answer": "\boxed{1}"
93
  },
94
  "ceb_Latn": {
95
  "lang_name_native": "Sinugbuanong Binisaya",
 
105
  "other_share": 0.0,
106
  "first_eng_pos": null,
107
  "top_other": null
108
+ },
109
+ "primer_user": "Siguruha nga kalma ang imong kamot kutob sa mahimo samtang padayon nga saktong maabot ang tanang nota - paningkamuti usab nga dili mobuhat og way hinungdang lihok gamit ang imong mga tudlo. Sa kani nga pamaagi, gamay ra ang imong kahago kutob sa maarangan. Hinumdumi nga dili kinahanglanng ig-on ang mga tekla gamit ang grabeng puwersa para sa dugang kakusgon sama sa piano. Sa akordyon, aron makakuha og dugang nga kakusog, kinahanglan gamiton nimo ang mga hasohasan nga adunay labing presyur o tulin.\n\nPangutana: Sumala sa istorya, unsay dili makonsiderar ingong tukmang tip para sa malamposong pagpatukar og akordyon?\n1. Para dugangan ang kakusog, patas-i ang puwersa sa pag-igo nimo sa mga tekla\n2. Likayi ang wala gikinahanglang mga lihok aron mapreserbar ang imong stamina\n3. Bantayi ang pag-igo sa mga nota samtang nagmentinar og kalma nga kamot\n4. Patas-i ang katulin sa imong pag-operate sa mga hasohasan aron makab-ot ang dugang kakusog\n\nTubaga sa numero sa hustong kapilian (1-4) sulod sa \\boxed{}.",
110
+ "primer_answer": "\boxed{1}"
111
  },
112
  "vie_Latn": {
113
  "lang_name_native": "Tiếng Việt",
 
123
  "other_share": 0.0,
124
  "first_eng_pos": null,
125
  "top_other": null
126
+ },
127
+ "primer_user": "Hãy đảm bảo tay của bạn được thư giãn nhất có thể khi nhấn tất cả các nốt nhạc chính xác - và cố gắng đừng làm nhiều cử động ngón tay không cần thiết. Chơi đàn theo cách này sẽ giúp bạn giảm mất sức tối đa. Hãy nhớ rằng, bạn không cần phải dùng nhiều lực để ấn phím nhằm tăng âm lượng như khi chơi đàn piano. Trên đàn phong cầm, để tăng âm lượng, bạn sử dụng ống thổi tác động mạnh hơn hoặc nhanh hơn.\n\nCâu hỏi: Theo đoạn văn, đâu không được coi là lời khuyên đúng để chơi được đàn phong cầm?\n1. Để tăng âm lượng, dùng thêm nhiều lực khi ấn phím\n2. Hạn chế tối thiểu động tác thừa để giữ sức\n3. Chú ý khi nhấn các nốt nhạc phải giữ cho tay được thư giãn nhất có thể\n4. Tăng tốc độ di chuyển hộp xếp để tăng âm lượng\n\nTrả lời bằng số của lựa chọn đúng (1-4) trong \\boxed{}.",
128
+ "primer_answer": "\boxed{1}"
129
  },
130
  "tha_Thai": {
131
  "lang_name_native": "ไทย",
 
141
  "other_share": 0.0,
142
  "first_eng_pos": null,
143
  "top_other": null
144
+ },
145
+ "primer_user": "ปล่อยให้มือของคุณผ่อนคลายที่สุดเท่าที่จะทำได้ขณะที่ยังคงเล่นโน้ตทุกตัวได้อย่างถูกต้อง พยายามไม่เคลื่อนไหวนิ้วมือมากเกินไปโดยไม่จำเป็น วิธีนี้จะทำให้ลดความเหน็ดเหนื่อยที่คุณได้รับให้เหลือน้อยที่สุด จำไว้ว่าไม่จำเป็นต้องกดแป้นให้แรงเพื่อให้เสียงดังขึ้นเหมือนเปียโน หากต้องการเพิ่มเสียงของหีบเพลงให้ดังขึ้น ให้ใช้แรงดันหรือความเร็วของที่สูบลมให้มากขึ้น\n\nคำถาม: จากบทความ ข้อใดไม่ใช่เคล็ดลับที่ถูกต้องสำหรับการเล่นหีบเพลงให้ประสบความสำเร็จ\n1. หากต้องการให้เสียงดังขึ้น ให้เพิ่มแรงดันตอนเล่นตัวโน้ต\n2. พยายามไม่เคลื่อนไหวนิ้วมือมากเกินไปเพื่อไม่ให้ออกแรงมากเกินไป\n3. ระวังตอนเล่นตัวโน้ตขณะที่ปล่อยให้มือผ่อนคลายที่สุดไปด้วย\n4. เพิ่มความเร็วของที่สูบลมเพื่อทำให้เสียงดังขึ้น\n\nตอบด้วยหมายเลขของตัวเลือกที่ถูกต้อง (1-4) ภายใน \\boxed{}.",
146
+ "primer_answer": "\boxed{1}"
147
  },
148
  "khm_Khmr": {
149
  "lang_name_native": "ខ្មែរ",
 
159
  "other_share": 0.0,
160
  "first_eng_pos": null,
161
  "top_other": null
162
+ },
163
+ "primer_user": "ត្រូវ​ឱ្យ​ប្រាកដ​ថា ដៃ​របស់​អ្នក​បាន​សម្រាក​តាម​ដែល​អាច​ធ្វើ​ទៅ​បាន នៅ​ខណៈ​ពេល​ដែល​អ្នក​អាច​នៅតែ​កត់ត្រា​បាន​ត្រឹមត្រូវ ម្យ៉ាងទៀត​ចូរ​កុំ​ព្យាយាម​ធ្វើ​ចលនា​ដោយ​ប្រើ​ម្រាម​ដៃ​របស់​អ្នក​បើ​មិន​ចាំបាច់។ វិធីនេះ​អ្នកនឹងធ្វើឱ្យខ្លួនឯងរាងហត់បន្តិច។ ចូរចងចាំថាមិន​ចាំបាច់​ចុច​តម្លើងសម្លេង​ដោយប្រើកម្លាំង​ខ្លាំង​ដូចនៅលើ​ព្យ៉ាណូទេ។ ដើម្បីបានកម្រិត​សំឡងពិសេសនៅលើអាគ័រឌីអុងនេះ ​អ្នកត្រូវប្រើប្រដាប់ផ្លុំខ្យល់​ដើម្បីបន្ថែម​សម្ពាធ ឬល្បឿន។\n\nសំណួរ: យោងតាមអត្ថបទសម្រង់នេះ តើអ្វីដែលមិនត្រូវបានចាត់ទុកថាជាគន្លឹះត្រឹមត្រូវសម្រាប់ការលេងអាគ័រដីយ៉ុងដោយជោគជ័យ?\n1. សម្រាប់កម្រិតសំឡេងបន្ថែម សូមបង្កើនកម្លាំងតាមដែលអ្នកសង្កត់គ្រាប់ចុច\n2. រក្សាចលនាដែលមិនចាំបាច់ជាអប្បបរមា ដើម្បីរក្សាភាពអំណត់របស់អ្នក\n3. ត្រូវចងចាំក្នុងសង្កត់ សញ្ញាភ្លេង ខណៈពេលទុកដៃឱ្យទំនេរ\n4. បង្កើនល្បឿនតាមរយៈរបស់ខាងក្រោម​ដែលអ្នកប្រើ​ ដើម្បីសម្រេចបានកម្រិតសំឡេងបន្ថែម\n\nឆ្លើយដោយលេខនៃជម្រើសត្រឹមត្រូវ (1-4) នៅក្នុង \\boxed{}.",
164
+ "primer_answer": "\boxed{1}"
165
  },
166
  "lao_Laoo": {
167
  "lang_name_native": "ພາສາລາວ",
 
177
  "other_share": 0.0,
178
  "first_eng_pos": null,
179
  "top_other": null
180
+ },
181
+ "primer_user": "ຕ້ອງໃຫ້ແນ່ໃຈວ່າມືຂອງທ່ານຜ່ອນຄາຍທີ່ສຸດເທົ່າທີ່ເປັນໄປໄດ້ໃນຂະນະທີ່ທ່ານຍັງກົດປຸ່ມບັນທຶກທັງໝົດຢ່າງຖືກຕ້ອງ ພ້ອມທັງພະຍາຍາມບໍ່ໃຫ້ນິ້ວມືຂອງທ່ານເຄື່ອນໄຫວຫຼາຍເກີນຂອບເຂດ. ວິທີນີ້, ທ່ານຈະເມື່ອຍກັບໂຕເອງໜ້ອຍທີ່ສຸດທີ່ຈະເປັນໄປໄດ້. ຈື່ໄວ້ວ່າບໍ່ຈຳເປັນຕ້ອງກົດປຸ່ມຕ່າງໆດ້ວຍແຮງຄ້າຍຄືກັບເປຍໂນ. ຕາມເງື່ອນໄຂແລ້ວ, ເພື່ອໃຫ້ໄດ້ປະລິມານຫຼາຍຂຶ້ນ ທ່ານຕ້ອງໃຊ້ບັ້ງສູບດ້ວຍແຮງດັນ ຫຼື ຄວາມໄວທີ່ຫຼາຍຂຶ້ນ.\n\nຄຳຖາມ: ອີງໃສ່ບົດອ່ານ, ສິ່ງໃດທີ່ບໍ່ຖືວ່າເປັນຄຳແນະນຳທີ່ຖືກຕ້ອງໃນການຫຼິ້ນແອັກຄໍດ້ຽນທີ່ສຳເລັດ?\n1. ເພື່ອເພີ່ມສຽງ, ເພີ່ມແຮງກົດຢ່າງເວລາທີ່ທ່ານກົດປຸມ\n2. ຮັກສາການເຄື່ອນໄຫວທີ່ບໍ່ຈຳເປັນໃຫ້ນ້ອຍທີ່ສຸດເພື່ອຮັກສາຄວາມແຂງແຮງຂອງທ່ານ\n3. ລະມັດລະວັງເມື່ອກົດສະໝຸດບັນທຶກໃນຂະນະທີ່ຮັກສາຄວາມຜ່ອນຄາຍຢູ່ມື\n4. ເພີ່ມຄວາມໄວໄດ້ເຊິ່ງທ່ານສາມາດດຳເນີນການປ້ຳເຄື່ອງສູບລົມເພື່ອໃຫ້ມີສຽງເພີ່ມຂຶ້ນ\n\nຕອບດ້ວຍເລກຂອງຕົວເລືອກທີ່ຖືກຕ້ອງ (1-4) ຢູ່ໃນ \\boxed{}.",
182
+ "primer_answer": "\boxed{1}"
183
  },
184
  "mya_Mymr": {
185
  "lang_name_native": "မြန်မာ",
 
195
  "other_share": 0.0,
196
  "first_eng_pos": null,
197
  "top_other": null
198
+ },
199
+ "primer_user": "ခလုတ်အားလုံးကို မှန်ကန်စွာတီးခတ်နေစဉ်တွင် သင်၏လက်ကို တတ်နိုင်သမျှ ဖြေလျှော့ထားပါ - ထို့အတူ လက်ချောင်းများကို မလိုအပ်ဘဲအလွန်အမင်းလှုပ်ရှားခြင်းမျိုးမလုပ်မိရန် ကြိုးစားပါ။ ထိုသို့သောနည်းလမ်းဖြင့်၊ သင်သည် ဖြည်းဖြည်းချင်း ပင်ပန်းလာပါလိမ့်မည်။ ဉပမာပြောရလျှင် စန္ဒယားတီးသည့်အခါ အသံကျယ်ကျယ်ထွက်ရှိရန်အတွက် ခလုတ်ကိုအားများများသုံးကာ ဖိတီးရန် မလိုသည်ကို သိရှိပါ။ အကော်ဒီယံတွင် အသံပိုရရန် အိတ်ကို ဖိအားသို့မဟုတ် အမြန်အသုံးပြုရသည်။\n\nမေးခွန်း: စာပိုဒ်အရ အကော်ဒီယံကိုအောင်မြင်စွာတီးခတ်ခြင်းအတွက် တိကျသော အကြံပြုချက်ဟု မယူဆနိုင်​သော အရာသည် အဘယ်နည်း။\n1. အသံကျယ်ကျယ်ထွက်ရှိရန်အတွက် ခလုတ်များကို အားပိုထည့်၍တီးခြင်း\n2. သင်၏သက်လုံကို ဆက်ထိန်းထားရန် မလိုအပ်ဘဲလှုပ်ရှားခြင်းမျိုးမလုပ်မိရန် ကြိုးစားခြင်း\n3. လက်ကိုဖြေလျှော့ထားလျက် ခလုတ်များတီးခတ်ခြင်းကို သတိထားခြင်း\n4. အသံပိုထွက်ရန် အိတ်က���ု ပို၍မြန်မြန်အသုံးပြုခြင်း\n\nမှန်ကန်သော ရွေးချယ်စရာ (1-4) ၏ နံပါတ်ကို \\boxed{} အတွင်း ဖြေပါ။",
200
+ "primer_answer": "\boxed{1}"
201
  },
202
  "hin_Deva": {
203
  "lang_name_native": "हिन्दी",
 
213
  "other_share": 0.0,
214
  "first_eng_pos": null,
215
  "top_other": null
216
+ },
217
+ "primer_user": "सुनिश्चित करें कि आपका हाथ सभी नोट्स को सही से बजाते हुए भी जितना हो सके, उतना स्थिर है - यह भी प्रयास करें कि आपकी उंगलियाँं अत्यधिक गतिवान न हो। इस तरह, आप जितना हो सकेगा, अपने को थका लेंगे. याद रखें कि ज़्यादा आवाज़ के लिए उसकी कुंजियों को ज़्यादा ज़ोर से दबाने की ज़रूरत नहीं होती है, जैसा कि पियानो में होता है. अकॉर्डियन पर, अतिरिक्त वॉल्यूम प्राप्त करने के लिए, आप कुहनी का उपयोग अधिक दबाव या गति के साथ करते हैं।\n\nप्रश्न: गद्यांश के अनुसार, संगीत के लिए अक्कादियन को सटीक रूप से पुनर्स्थापित करने के लिए वास्तव में क्या संभव नहीं है?\n1. अधिक मात्रा में उत्पन्न करने के लिए कुंजियों को अधिक बल के साथ दबाया जाना चाहिए\n2. अपनी उंगलियों से अनावश्यक हरकत करने से बचें ताकि आपकी शक्ति कम न हो\n3. यह सुनिश्चित करने के लिए ध्यान रखा जाना चाहिए कि आपके हाथ अक्काडियन को सही ढंग से पुनः प्राप्त करने के लिए सहज हों\n4. उच्च मात्रा प्राप्त करने के लिए उच्च दबाव या वेग वाले अकॉर्डियन का उपयोग किया जा सकता है\n\nसही विकल्प का नंबर (1-4) \\boxed{} के अंदर दें।",
218
+ "primer_answer": "\boxed{1}"
219
  },
220
  "tam_Taml": {
221
  "lang_name_native": "தமிழ்",
 
231
  "other_share": 0.0,
232
  "first_eng_pos": null,
233
  "top_other": null
234
+ },
235
+ "primer_user": "அனைத்து சுரங்களையும் சரியாக வாசிக்கும் போது உன் கை முடிந்தவரை சௌகரியமாக இருப்பதை உறுதி செய்து கொள்ளவும், மேலும், விரல்களால் அதிக தேவையற்ற அசைவுகளை செய்யாமல் இருக்க முயலவும். இந்த வழியில், முடிந்த அளவு குறைவாக சோர்வடைவீர்கள். அதிக ஒலிக்காக விசைகளை பியானோவில் அழுத்துவது போல அதிக ஆற்றலுடன் அழுத்தத் தேவையில்லை என்பதை நினைவில் கொள்ளுங்கள். அக்கார்டியனில் அதிக ஒலியினைப் பெறுவதற்கு, நீங்கள் அதிக அழுத்தம் அல்லது வேகமுடைய துருத்திகளைப் பயன்படுத்தலாம்.\n\nகேள்வி: கொடுக்கப்பட்டிருக்கும் இந்த பத்தியில், சுரங்களை துல்லியமாக வாசிப்பதற்கான சரியான உதவிக் குறிப்பு இல்லை எனக் கருதப்படுவது எது?\n1. அதிக அளவு ஒலியினை ஏற்படுத்த, விசைகளை அதிக அளவு ஆற்றலுடன் அழுத்த வேண்டும்\n2. உங்கள் ஸ்டெமினா குறையாமல் இருக்க விரல்களால் தேவையற்ற அசைவுகளைச் செய்யாமல் இருக்க வேண்டும்\n3. சுரங்களை சரியாக வாசிக்கும் போது உங்கள் கைகளை சௌகரியமாக வைத்திருக்கிறீர்களா என்பதில் கவனம் செலுத்த வேண்டும்\n4. அதிக அளவு ஒலியினைப் பெறுவதற்கு, அதிக அழுத்தத்தையோ அல்லது வேகத்தையோ கொண்ட துருத்திகளைப் பயன்படுத்தலாம்\n\nசரியான விருப்பத்தின் எண்ணை (1-4) \\boxed{} உள்ளே கொடுக்கவும்.",
236
+ "primer_answer": "\boxed{1}"
237
  },
238
  "kor_Hang": {
239
  "lang_name_native": "한국어",
 
249
  "other_share": 0.0,
250
  "first_eng_pos": null,
251
  "top_other": null
252
+ },
253
+ "primer_user": "모든 음을 올바르게 누른 상태에서 손을 최대한 편하게 둡니다. 또한 손가락으로 관련 없는 동작을 많이 하지 않도록 하십시오. 이 방식을 통해 연주의 피로도를 최대한 낮출 수 있습니다. 피아노와 달리, 아코디언은 더 큰 소리를 내기 위해 건반을 세게 칠 필요가 없다는 것을 기억하세요. 아코디언에 볼륨을 더 주려면 풀무에 힘을 더 주거나 빠르게 풀무를 사용해야 합니다.\n\n질문: 지문에 따르면, 아코디언을 잘 연주하기 위한 정확한 팁으로 볼 수 없는 것은 무엇입니까?\n1. 볼륨을 높이려면 키를 더 세게 누른다\n2. 체력을 유지하기 위해 불필요한 움직임을 최소화한다\n3. 손을 최대한 편하게 유지하면서 정확한 음표를 치는 데 신경을 쓴다\n4. 볼륨을 추가로 높이려면 풀무를 작동하는 속도를 높인다\n\n정답 선택지의 번호(1-4)를 \\boxed{} 안에 입력하세요.",
254
+ "primer_answer": "\boxed{1}"
255
  },
256
  "jpn_Jpan": {
257
  "lang_name_native": "日本語",
 
267
  "other_share": 0.0,
268
  "first_eng_pos": null,
269
  "top_other": null
270
+ },
271
+ "primer_user": "音符どおりに正確に奏でながら、手をできるだけリラックスした状態で保つようにしましょう。また、指が余計な動きをしないように努めましょう。 そうすることで、疲れがたまりにくくなります。ピアノのように音量を出すために力を入れて鍵盤を叩く必要はないことを覚えておいてください。 アコーディオンを弾くときに音量を上げるには、蛇腹を強く、または速く動かします。\n\n質問: 前述の内容によると、上手にアコーディオンを弾くための適切なコツとして考えられないものはどれでしょう?\n1. 音量を上げるために、鍵盤を叩く力を強くする\n2. 体力を温存するために、無駄な動きを最小限に抑える\n3. 手をリラックスした状態に保ちながら、音符通りに弾くように注意する\n4. 音量を上げるために蛇腹を動かす速度を上げる\n\n正解の選択肢の番号(1-4)を\\boxed{}の中に入れて答えてください。",
272
+ "primer_answer": "\boxed{1}"
273
  },
274
  "arb_Arab": {
275
  "lang_name_native": "العربية",
 
285
  "other_share": 0.0,
286
  "first_eng_pos": null,
287
  "top_other": null
288
+ },
289
+ "primer_user": "تأكد من استرخاء يدك قدر الإمكان مع الاستمرار في ضرب كل النغمات بشكل صحيح - حاول كذلك عدم القيام بحركاتٍ غريبةٍ بأصابعك. لن تبذل مجهوداً كبيراً إذا اتبعت تلك الطريقة. ضع نصب عينيك أنه ليس عليك الضغط على مفاتيح الأكورديون بقوةٍ كبيرةٍ لتحصل على مستوى صوتٍ أعلى كما في البيانو. في الأكورديون، للحصول على صوت أعلى إضافي، يمكنك استخدام المنفاخ بمزيد من الضغط أو السرعة.\n\nسؤال: وفقاً للفقرة، ما الذي لا يُعتبر نصيحة دقيقة للعزف الناجح على الأكورديون؟\n1. للحصول على صوت أعلى، زدّ القوة التي تضغط بها على المفاتيح\n2. أبقِ الحركة غير الضرورية في حدها الأدنى للحفاظ على طاقتك\n3. احرص على ضرب النغمات مع إبقاء اليد مسترخية\n4. زدّ السرعة التي تُشغل بها المنفاخ للحصول على صوت إضافي\n\nأجب برقم الخيار الصحيح (1-4) داخل \\boxed{}.",
290
+ "primer_answer": "\boxed{1}"
291
  },
292
  "yor_Latn": {
293
  "lang_name_native": "Yorùbá",
 
303
  "other_share": 0.0,
304
  "first_eng_pos": null,
305
  "top_other": null
306
+ },
307
+ "primer_user": "Rí i dájú pé ọwọ́ rẹ ní ìsinmi bó se yẹ nígbàtí o bá ń kọlu gbogbo àkíyèsí tí ó tó - pẹ̀lú pé ki o má dààmú ọmọ ìka ọwọ́ rẹ jùlọ. Bó ba n ṣe báyí kò ní rẹ̀ ẹ́ jù. Rántí pé o kò níláti fi agbára tẹ ojú rẹ̀ kí ohun rẹ̀ ga síi bí ti piano. Láti mú ohun akodioni tóbi si wàá lò àwọn belò pẹ̀lú agbára tàbí ere sí i.\n\nÌbéèrè: Gẹgẹ bi oju ewe naa, kini a o pe ni ojulowo itọni fun imọ titẹ àkọdiọnu daadaa??\n1. Fun afikun iwọn didun, lo agbara pipọ pẹlu eyiti o lu awọn bọtini naa\n2. Maṣe maa rin ainidi kiri ki o le ni okun si\n3. Ṣe akiyesi lilu awọn nooti lakoko mimu ọwọ rẹ rọ\n4. Mu iyara pọ si pẹlu eyi ti o fi ṣẹ awọn ohun agbara isalẹ lati ṣaṣeyọri iwọn didun afikun\n\nDáhùn pẹ̀lú nọ́mbà àṣàyàn tó tọ́ (1-4) nínú \\boxed{}.",
308
+ "primer_answer": "\boxed{1}"
309
  },
310
  "plt_Latn": {
311
  "lang_name_native": "fiteny malagasy",
 
321
  "other_share": 0.0,
322
  "first_eng_pos": null,
323
  "top_other": null
324
+ },
325
+ "primer_user": "Ataovy azo antoka fa milamina tsara ny tanana eo am-panoratana ny naoty araka ny tokony ho izy - ary koa miezaha tsy dia manao fihetsika be amin’ny ratsan-tananao. Amin’ity fomba ity dia mba tsy ho reraka be ny tenanao. Tadidio fa tsy ilaina ny manindry ny bokotra mafy mba hahazoana feo mafy toy ny amin’ny piano. Amin’ny angorodao, ahazoana feo mafy tsara, dia tsindrina mafimafy ary haingana kokoa ny lakolosy.\n\nFanontaniana: Araka ilay sombin-dahatsoratra, inona no tsy heverina ho torohevitra mitombina hitendrena angorodao am-pahombiazana?\n1. Hampiakarana ny feony, ampitomboy ny hery amin’ny fanindrinao ny bokotra\n2. Ahenao ny fihetsika tsy ilaina mba hitazomana ny tanjakao\n3. Mifantoha lalina amin’ny naoty sady hazony mba ho malefaka hatrany ny tanana\n4. Ampitomboy ny hafainganana amin’ny tafoforana mba ahatrarana ny feo avo indrindra\n\nValio amin'ny laharan'ny safidy marina (1-4) ao anatin'ny \\boxed{}.",
326
+ "primer_answer": "\boxed{1}"
327
  }
328
  }
data/sea_thinking.jsonl CHANGED
The diff for this file is too large to render. See raw diff
 
results.json CHANGED
@@ -835,206 +835,294 @@
835
  "empty_or_truncated": 0,
836
  "reasoning_tokens": 471.4,
837
  "completion_tokens": 482.7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
838
  }
839
  },
840
  "math": {
841
- "ceb_Latn|C0|low": {
842
- "n": 51,
843
- "acc": 98.0,
844
- "target_share": 0.175,
845
- "eng_share": 0.742,
846
- "pct_mostly_target": 5.9,
847
  "no_reasoning_signal": 0,
848
  "empty_or_truncated": 0,
849
- "reasoning_tokens": 584.588,
850
- "completion_tokens": 664.294
 
 
 
 
 
 
 
 
 
 
 
851
  },
852
  "eng_Latn|C0|low": {
853
- "n": 52,
854
- "acc": 94.2,
855
- "target_share": 0.923,
856
  "eng_share": 0.0,
857
  "pct_mostly_target": 100.0,
858
- "no_reasoning_signal": 0,
859
  "empty_or_truncated": 0,
860
- "reasoning_tokens": 529.962,
861
- "completion_tokens": 579.423
862
  },
863
- "ind_Latn|C2|low": {
864
- "n": 49,
865
- "acc": 98.0,
866
- "target_share": 0.681,
867
- "eng_share": 0.22,
868
- "pct_mostly_target": 69.4,
869
- "no_reasoning_signal": 0,
870
  "empty_or_truncated": 0,
871
- "reasoning_tokens": 507.755,
872
- "completion_tokens": 581.878
873
  },
874
- "zsm_Latn|C0|low": {
875
- "n": 51,
876
- "acc": 90.2,
877
- "target_share": 0.221,
878
- "eng_share": 0.733,
879
- "pct_mostly_target": 11.8,
880
  "no_reasoning_signal": 0,
881
  "empty_or_truncated": 0,
882
- "reasoning_tokens": 685.863,
883
- "completion_tokens": 740.549
884
  },
885
- "vie_Latn|C2|low": {
886
- "n": 49,
887
- "acc": 87.8,
888
- "target_share": 0.512,
889
- "eng_share": 0.432,
890
- "pct_mostly_target": 46.9,
891
  "no_reasoning_signal": 0,
892
  "empty_or_truncated": 0,
893
- "reasoning_tokens": 802.714,
894
- "completion_tokens": 876.367
895
  },
896
- "ceb_Latn|C2|low": {
897
- "n": 48,
898
- "acc": 93.8,
899
- "target_share": 0.228,
900
- "eng_share": 0.683,
901
- "pct_mostly_target": 10.4,
902
  "no_reasoning_signal": 0,
903
  "empty_or_truncated": 0,
904
- "reasoning_tokens": 572.667,
905
- "completion_tokens": 711.812
906
  },
907
- "jav_Latn|C2|low": {
908
- "n": 50,
909
- "acc": 96.0,
910
- "target_share": 0.145,
911
- "eng_share": 0.554,
912
- "pct_mostly_target": 0.0,
913
  "no_reasoning_signal": 0,
914
  "empty_or_truncated": 0,
915
- "reasoning_tokens": 582.74,
916
- "completion_tokens": 761.28
917
  },
918
- "ind_Latn|C0|low": {
919
- "n": 49,
920
- "acc": 98.0,
921
- "target_share": 0.217,
922
- "eng_share": 0.718,
923
- "pct_mostly_target": 12.2,
924
  "no_reasoning_signal": 0,
925
  "empty_or_truncated": 0,
926
- "reasoning_tokens": 517.878,
927
- "completion_tokens": 558.612
928
  },
929
  "zsm_Latn|C2|low": {
930
- "n": 50,
931
- "acc": 96.0,
932
- "target_share": 0.579,
933
- "eng_share": 0.372,
934
- "pct_mostly_target": 52.0,
935
  "no_reasoning_signal": 0,
936
  "empty_or_truncated": 0,
937
- "reasoning_tokens": 683.08,
938
- "completion_tokens": 827.5
939
  },
940
  "tgl_Latn|C0|low": {
941
- "n": 51,
942
- "acc": 98.0,
943
- "target_share": 0.182,
944
- "eng_share": 0.74,
945
- "pct_mostly_target": 2.0,
946
  "no_reasoning_signal": 0,
947
  "empty_or_truncated": 0,
948
- "reasoning_tokens": 610.49,
949
- "completion_tokens": 684.51
950
  },
951
- "eng_Latn|C2|low": {
952
- "n": 52,
953
- "acc": 94.2,
954
- "target_share": 0.92,
955
- "eng_share": 0.0,
956
- "pct_mostly_target": 100.0,
957
  "no_reasoning_signal": 0,
958
  "empty_or_truncated": 0,
959
- "reasoning_tokens": 551.577,
960
- "completion_tokens": 605.135
961
  },
962
- "sun_Latn|C2|low": {
963
- "n": 51,
964
- "acc": 98.0,
965
- "target_share": 0.167,
966
- "eng_share": 0.558,
967
- "pct_mostly_target": 2.0,
968
  "no_reasoning_signal": 0,
969
  "empty_or_truncated": 0,
970
- "reasoning_tokens": 499.353,
971
- "completion_tokens": 706.51
972
  },
973
- "jav_Latn|C0|low": {
974
- "n": 46,
975
- "acc": 95.7,
976
- "target_share": 0.172,
977
- "eng_share": 0.715,
978
- "pct_mostly_target": 2.2,
979
  "no_reasoning_signal": 0,
980
  "empty_or_truncated": 0,
981
- "reasoning_tokens": 570.609,
982
- "completion_tokens": 647.674
983
  },
984
- "tgl_Latn|C2|low": {
985
- "n": 49,
986
- "acc": 95.9,
987
- "target_share": 0.249,
988
- "eng_share": 0.679,
989
- "pct_mostly_target": 16.3,
 
 
 
 
 
 
 
 
 
 
 
990
  "no_reasoning_signal": 0,
991
  "empty_or_truncated": 0,
992
- "reasoning_tokens": 608,
993
- "completion_tokens": 834.122
994
  },
995
- "vie_Latn|C0|low": {
996
- "n": 51,
997
- "acc": 88.2,
998
- "target_share": 0.211,
999
- "eng_share": 0.735,
1000
- "pct_mostly_target": 11.8,
1001
  "no_reasoning_signal": 0,
1002
  "empty_or_truncated": 0,
1003
- "reasoning_tokens": 827.392,
1004
- "completion_tokens": 860.137
1005
  },
1006
- "sun_Latn|C0|low": {
1007
- "n": 48,
1008
- "acc": 97.9,
1009
- "target_share": 0.184,
1010
- "eng_share": 0.706,
1011
- "pct_mostly_target": 4.2,
1012
  "no_reasoning_signal": 0,
1013
  "empty_or_truncated": 0,
1014
- "reasoning_tokens": 547.312,
1015
- "completion_tokens": 581.562
1016
  },
1017
- "tha_Thai|C0|low": {
1018
- "n": 54,
1019
- "acc": 92.6,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1020
  "target_share": 0.131,
1021
- "eng_share": 0.808,
1022
- "pct_mostly_target": 1.9,
1023
- "no_reasoning_signal": 0,
1024
  "empty_or_truncated": 0,
1025
- "reasoning_tokens": 740.704,
1026
- "completion_tokens": 762.019
1027
  },
1028
- "tha_Thai|C2|low": {
1029
- "n": 52,
1030
- "acc": 94.2,
1031
- "target_share": 0.416,
1032
- "eng_share": 0.523,
1033
- "pct_mostly_target": 34.6,
1034
  "no_reasoning_signal": 0,
1035
  "empty_or_truncated": 0,
1036
- "reasoning_tokens": 835.231,
1037
- "completion_tokens": 913.731
1038
  }
1039
  },
1040
  "replication": [
@@ -1475,7 +1563,39 @@
1475
  "first_eng_pos": 0.0,
1476
  "top_other": "ind_Latn"
1477
  },
1478
- "K3/prompt_tokens": 402849
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1479
  },
1480
  "lid_calibration": {
1481
  "eng_Latn": {
@@ -1800,9 +1920,13 @@
1800
  "pass": true,
1801
  "detail": "C3 > C2 in 13/18"
1802
  },
 
 
 
 
1803
  "H5": {
1804
  "pass": true,
1805
- "detail": "mean acc(C2)-acc(C0) = 0.12 pts over 9 langs"
1806
  },
1807
  "H6": {
1808
  "pass": false,
@@ -1811,10 +1935,6 @@
1811
  "H7": {
1812
  "pass": true,
1813
  "detail": "K3 think target_share = 0.438"
1814
- },
1815
- "H4": {
1816
- "pass": true,
1817
- "detail": "Spearman(Joshi tier, C2 target_share) = 0.281 over 18 languages (pre-registration phrased it as a negative correlation with 'low-resource-ness'; tier is ordered high=5)"
1818
  }
1819
  },
1820
  "atlas_eng_share_ci95": {
@@ -2685,14 +2805,14 @@
2685
  }
2686
  },
2687
  "math_mcnemar_C0_vs_C2": {
2688
- "C0_only_correct": 4,
2689
  "C2_only_correct": 6,
2690
- "p_exact": 0.7539
2691
  },
2692
  "routes": {
2693
- "reflection.ai": 1987,
2694
- "huggingface": 1053
2695
  },
2696
- "cost_usd": 2.6405,
2697
- "sea_thinking_rows": 174
2698
  }
 
835
  "empty_or_truncated": 0,
836
  "reasoning_tokens": 471.4,
837
  "completion_tokens": 482.7
838
+ },
839
+ "ind_Latn|C0|high": {
840
+ "n": 40,
841
+ "acc": 95.0,
842
+ "target_share": 0.417,
843
+ "eng_share": 0.565,
844
+ "pct_mostly_target": 37.5,
845
+ "no_reasoning_signal": 0,
846
+ "empty_or_truncated": 0,
847
+ "reasoning_tokens": 487.975,
848
+ "completion_tokens": 503.9
849
+ },
850
+ "ind_Latn|C2|high": {
851
+ "n": 40,
852
+ "acc": 95.0,
853
+ "target_share": 0.784,
854
+ "eng_share": 0.202,
855
+ "pct_mostly_target": 82.5,
856
+ "no_reasoning_signal": 0,
857
+ "empty_or_truncated": 0,
858
+ "reasoning_tokens": 366.575,
859
+ "completion_tokens": 384.475
860
+ },
861
+ "jav_Latn|C0|high": {
862
+ "n": 40,
863
+ "acc": 90.0,
864
+ "target_share": 0.49,
865
+ "eng_share": 0.459,
866
+ "pct_mostly_target": 45.0,
867
+ "no_reasoning_signal": 0,
868
+ "empty_or_truncated": 0,
869
+ "reasoning_tokens": 750.975,
870
+ "completion_tokens": 765.925
871
+ },
872
+ "jav_Latn|C2|high": {
873
+ "n": 40,
874
+ "acc": 95.0,
875
+ "target_share": 0.423,
876
+ "eng_share": 0.531,
877
+ "pct_mostly_target": 32.5,
878
+ "no_reasoning_signal": 0,
879
+ "empty_or_truncated": 0,
880
+ "reasoning_tokens": 656.05,
881
+ "completion_tokens": 726.025
882
  }
883
  },
884
  "math": {
885
+ "ind_Latn|C0|low": {
886
+ "n": 58,
887
+ "acc": 96.6,
888
+ "target_share": 0.233,
889
+ "eng_share": 0.697,
890
+ "pct_mostly_target": 15.5,
891
  "no_reasoning_signal": 0,
892
  "empty_or_truncated": 0,
893
+ "reasoning_tokens": 485.224,
894
+ "completion_tokens": 522.931
895
+ },
896
+ "ind_Latn|C2|low": {
897
+ "n": 58,
898
+ "acc": 96.6,
899
+ "target_share": 0.713,
900
+ "eng_share": 0.197,
901
+ "pct_mostly_target": 73.7,
902
+ "no_reasoning_signal": 1,
903
+ "empty_or_truncated": 0,
904
+ "reasoning_tokens": 467.914,
905
+ "completion_tokens": 535.81
906
  },
907
  "eng_Latn|C0|low": {
908
+ "n": 58,
909
+ "acc": 94.8,
910
+ "target_share": 0.926,
911
  "eng_share": 0.0,
912
  "pct_mostly_target": 100.0,
913
+ "no_reasoning_signal": 1,
914
  "empty_or_truncated": 0,
915
+ "reasoning_tokens": 497.81,
916
+ "completion_tokens": 543.362
917
  },
918
+ "eng_Latn|C2|low": {
919
+ "n": 58,
920
+ "acc": 94.8,
921
+ "target_share": 0.923,
922
+ "eng_share": 0.0,
923
+ "pct_mostly_target": 100.0,
924
+ "no_reasoning_signal": 1,
925
  "empty_or_truncated": 0,
926
+ "reasoning_tokens": 531.948,
927
+ "completion_tokens": 581.103
928
  },
929
+ "jav_Latn|C0|low": {
930
+ "n": 58,
931
+ "acc": 96.6,
932
+ "target_share": 0.147,
933
+ "eng_share": 0.702,
934
+ "pct_mostly_target": 1.7,
935
  "no_reasoning_signal": 0,
936
  "empty_or_truncated": 0,
937
+ "reasoning_tokens": 531.862,
938
+ "completion_tokens": 595.586
939
  },
940
+ "jav_Latn|C2|low": {
941
+ "n": 58,
942
+ "acc": 96.6,
943
+ "target_share": 0.146,
944
+ "eng_share": 0.549,
945
+ "pct_mostly_target": 1.7,
946
  "no_reasoning_signal": 0,
947
  "empty_or_truncated": 0,
948
+ "reasoning_tokens": 555.741,
949
+ "completion_tokens": 722.793
950
  },
951
+ "sun_Latn|C0|low": {
952
+ "n": 58,
953
+ "acc": 96.6,
954
+ "target_share": 0.184,
955
+ "eng_share": 0.688,
956
+ "pct_mostly_target": 6.9,
957
  "no_reasoning_signal": 0,
958
  "empty_or_truncated": 0,
959
+ "reasoning_tokens": 526.207,
960
+ "completion_tokens": 556.586
961
  },
962
+ "sun_Latn|C2|low": {
963
+ "n": 58,
964
+ "acc": 94.8,
965
+ "target_share": 0.165,
966
+ "eng_share": 0.556,
967
+ "pct_mostly_target": 1.7,
968
  "no_reasoning_signal": 0,
969
  "empty_or_truncated": 0,
970
+ "reasoning_tokens": 513.983,
971
+ "completion_tokens": 701.707
972
  },
973
+ "zsm_Latn|C0|low": {
974
+ "n": 58,
975
+ "acc": 91.4,
976
+ "target_share": 0.206,
977
+ "eng_share": 0.726,
978
+ "pct_mostly_target": 12.1,
979
  "no_reasoning_signal": 0,
980
  "empty_or_truncated": 0,
981
+ "reasoning_tokens": 653.483,
982
+ "completion_tokens": 703.017
983
  },
984
  "zsm_Latn|C2|low": {
985
+ "n": 58,
986
+ "acc": 94.8,
987
+ "target_share": 0.603,
988
+ "eng_share": 0.345,
989
+ "pct_mostly_target": 55.2,
990
  "no_reasoning_signal": 0,
991
  "empty_or_truncated": 0,
992
+ "reasoning_tokens": 623.155,
993
+ "completion_tokens": 757.845
994
  },
995
  "tgl_Latn|C0|low": {
996
+ "n": 58,
997
+ "acc": 98.3,
998
+ "target_share": 0.167,
999
+ "eng_share": 0.756,
1000
+ "pct_mostly_target": 1.7,
1001
  "no_reasoning_signal": 0,
1002
  "empty_or_truncated": 0,
1003
+ "reasoning_tokens": 579.879,
1004
+ "completion_tokens": 646.397
1005
  },
1006
+ "tgl_Latn|C2|low": {
1007
+ "n": 58,
1008
+ "acc": 96.6,
1009
+ "target_share": 0.228,
1010
+ "eng_share": 0.691,
1011
+ "pct_mostly_target": 15.5,
1012
  "no_reasoning_signal": 0,
1013
  "empty_or_truncated": 0,
1014
+ "reasoning_tokens": 561.707,
1015
+ "completion_tokens": 776.224
1016
  },
1017
+ "ceb_Latn|C0|low": {
1018
+ "n": 58,
1019
+ "acc": 94.8,
1020
+ "target_share": 0.181,
1021
+ "eng_share": 0.732,
1022
+ "pct_mostly_target": 8.6,
1023
  "no_reasoning_signal": 0,
1024
  "empty_or_truncated": 0,
1025
+ "reasoning_tokens": 587.241,
1026
+ "completion_tokens": 659.828
1027
  },
1028
+ "ceb_Latn|C2|low": {
1029
+ "n": 58,
1030
+ "acc": 91.4,
1031
+ "target_share": 0.216,
1032
+ "eng_share": 0.69,
1033
+ "pct_mostly_target": 12.1,
1034
  "no_reasoning_signal": 0,
1035
  "empty_or_truncated": 0,
1036
+ "reasoning_tokens": 560.5,
1037
+ "completion_tokens": 684.603
1038
  },
1039
+ "vie_Latn|C0|low": {
1040
+ "n": 58,
1041
+ "acc": 87.9,
1042
+ "target_share": 0.213,
1043
+ "eng_share": 0.736,
1044
+ "pct_mostly_target": 12.5,
1045
+ "no_reasoning_signal": 2,
1046
+ "empty_or_truncated": 0,
1047
+ "reasoning_tokens": 755.207,
1048
+ "completion_tokens": 785.448
1049
+ },
1050
+ "vie_Latn|C2|low": {
1051
+ "n": 58,
1052
+ "acc": 87.9,
1053
+ "target_share": 0.561,
1054
+ "eng_share": 0.388,
1055
+ "pct_mostly_target": 53.4,
1056
  "no_reasoning_signal": 0,
1057
  "empty_or_truncated": 0,
1058
+ "reasoning_tokens": 732.362,
1059
+ "completion_tokens": 796.397
1060
  },
1061
+ "tha_Thai|C0|low": {
1062
+ "n": 58,
1063
+ "acc": 93.1,
1064
+ "target_share": 0.129,
1065
+ "eng_share": 0.811,
1066
+ "pct_mostly_target": 1.7,
1067
  "no_reasoning_signal": 0,
1068
  "empty_or_truncated": 0,
1069
+ "reasoning_tokens": 714.224,
1070
+ "completion_tokens": 734.655
1071
  },
1072
+ "tha_Thai|C2|low": {
1073
+ "n": 58,
1074
+ "acc": 94.8,
1075
+ "target_share": 0.438,
1076
+ "eng_share": 0.506,
1077
+ "pct_mostly_target": 37.9,
1078
  "no_reasoning_signal": 0,
1079
  "empty_or_truncated": 0,
1080
+ "reasoning_tokens": 792.776,
1081
+ "completion_tokens": 864.655
1082
  },
1083
+ "ind_Latn|C3|low": {
1084
+ "n": 58,
1085
+ "acc": 94.8,
1086
+ "target_share": 0.666,
1087
+ "eng_share": 0.245,
1088
+ "pct_mostly_target": 68.5,
1089
+ "no_reasoning_signal": 4,
1090
+ "empty_or_truncated": 0,
1091
+ "reasoning_tokens": 358.172,
1092
+ "completion_tokens": 370.948
1093
+ },
1094
+ "jav_Latn|C3|low": {
1095
+ "n": 58,
1096
+ "acc": 93.1,
1097
+ "target_share": 0.287,
1098
+ "eng_share": 0.507,
1099
+ "pct_mostly_target": 26.3,
1100
+ "no_reasoning_signal": 1,
1101
+ "empty_or_truncated": 0,
1102
+ "reasoning_tokens": 419.897,
1103
+ "completion_tokens": 435.345
1104
+ },
1105
+ "sun_Latn|C3|low": {
1106
+ "n": 58,
1107
+ "acc": 94.8,
1108
  "target_share": 0.131,
1109
+ "eng_share": 0.688,
1110
+ "pct_mostly_target": 7.0,
1111
+ "no_reasoning_signal": 1,
1112
  "empty_or_truncated": 0,
1113
+ "reasoning_tokens": 421.552,
1114
+ "completion_tokens": 440.724
1115
  },
1116
+ "tha_Thai|C3|low": {
1117
+ "n": 58,
1118
+ "acc": 93.1,
1119
+ "target_share": 0.262,
1120
+ "eng_share": 0.667,
1121
+ "pct_mostly_target": 24.1,
1122
  "no_reasoning_signal": 0,
1123
  "empty_or_truncated": 0,
1124
+ "reasoning_tokens": 539,
1125
+ "completion_tokens": 551.862
1126
  }
1127
  },
1128
  "replication": [
 
1563
  "first_eng_pos": 0.0,
1564
  "top_other": "ind_Latn"
1565
  },
1566
+ "K3/prompt_tokens": 402849,
1567
+ "K0/think_noquote": {
1568
+ "n_seg": 40,
1569
+ "target_share": 0.0,
1570
+ "eng_share": 0.313,
1571
+ "other_share": 0.687,
1572
+ "first_eng_pos": 0.0,
1573
+ "top_other": "ind_Latn"
1574
+ },
1575
+ "K3/think_noquote": {
1576
+ "n_seg": 151,
1577
+ "target_share": 0.067,
1578
+ "eng_share": 0.052,
1579
+ "other_share": 0.881,
1580
+ "first_eng_pos": 0.0,
1581
+ "top_other": "ind_Latn"
1582
+ },
1583
+ "K2/think_noquote": {
1584
+ "n_seg": 115,
1585
+ "target_share": 0.253,
1586
+ "eng_share": 0.127,
1587
+ "other_share": 0.62,
1588
+ "first_eng_pos": 0.0,
1589
+ "top_other": "ind_Latn"
1590
+ },
1591
+ "K1/think_noquote": {
1592
+ "n_seg": 76,
1593
+ "target_share": 0.221,
1594
+ "eng_share": 0.188,
1595
+ "other_share": 0.591,
1596
+ "first_eng_pos": 0.0,
1597
+ "top_other": "ind_Latn"
1598
+ }
1599
  },
1600
  "lid_calibration": {
1601
  "eng_Latn": {
 
1920
  "pass": true,
1921
  "detail": "C3 > C2 in 13/18"
1922
  },
1923
+ "H4": {
1924
+ "pass": true,
1925
+ "detail": "Spearman(Joshi tier, C2 target_share) = 0.281 over 18 languages (pre-registration phrased it as a negative correlation with 'low-resource-ness'; tier is ordered high=5); permutation p = 0.269 (direction as predicted, not significant)"
1926
+ },
1927
  "H5": {
1928
  "pass": true,
1929
+ "detail": "mean acc(C2)-acc(C0) = -0.2 pts over 9 langs"
1930
  },
1931
  "H6": {
1932
  "pass": false,
 
1935
  "H7": {
1936
  "pass": true,
1937
  "detail": "K3 think target_share = 0.438"
 
 
 
 
1938
  }
1939
  },
1940
  "atlas_eng_share_ci95": {
 
2805
  }
2806
  },
2807
  "math_mcnemar_C0_vs_C2": {
2808
+ "C0_only_correct": 7,
2809
  "C2_only_correct": 6,
2810
+ "p_exact": 1.0
2811
  },
2812
  "routes": {
2813
+ "reflection.ai": 2360,
2814
+ "huggingface": 2116
2815
  },
2816
+ "cost_usd": 2.8987,
2817
+ "sea_thinking_rows": 1453
2818
  }