barandinho/turkish-mcq-rlvr

Turkish MCQ Reasoning Dataset for RLVR training

What this dataset is

This dataset is a Turkish multiple-choice (MCQ) reasoning benchmark augmented with a weak-model pass-rate difficulty signal, designed for curriculum learning, GRPO / RLVR-style training, and evaluation of reasoning models on Turkish exam-style questions.

It is built from a Turkish MMLU-style MCQ corpus and annotated with a model-based difficulty proxy (pass_rate), computed via repeated sampling of a weak baseline model and automatic answer parsing.

Each example represents one Turkish MCQ question (question + options + correct choice), together with a difficulty estimate based on how often a weak model answers it correctly among the runs where the answer could be parsed.

---

Data source

This dataset is originally based on:

• alibayram/turkish_mmlu

It preserves the original subject structure (bolum, konu) and MCQ fields (soru, secenekler, cevap) while adding an RL-friendly weak-model difficulty signal.

---

Pass rate: what it means

For each question:

• A weak Turkish-capable baseline model (Qwen3-4B no-thinking variant) is sampled 4 independent times
• Same prompt, stochastic decoding
• Each generation is passed through an answer parsing function (to extract the chosen option)
• If a generation cannot be parsed, it is excluded from scoring (it does not count as wrong — it counts as “not an attempt”)

Key detail: “attempt” means parsable output

In this dataset:

• Total samples per question: 4
• Attempts: number of samples (out of 4) where the model output is parsable into a valid MCQ choice (attempted ∈ {0, 1, 2, 3, 4})

Correctness is only measured over attempted parses.

---

pass_rate definition

Let:

• attempted = number of parsable outputs (0–4)
• correct = number of parsable outputs that match the gold answer

Then:

pass_rate = correct / attempted

If attempted == 0, pass_rate is defined as 0.0.

Because attempted can vary per question (due to parse failures), pass_rate can take values like:

• 0.0
• 0.25, 0.5, 0.75 (when attempted = 4)
• 0.333..., 0.666... (when attempted = 3)
• 1.0 (filtered out in this release)

Interpretation

pass_rate	Interpretation (weak model)
0.00	Never correct (or never parsable) → very hard / noisy
0.25–0.50	Sometimes correct → hard / medium
0.66–0.75	Often correct → medium / easier
1.00	Always correct → too easy (filtered out)

This yields a weak-to-strong difficulty signal without human difficulty labels, useful for:

• Curriculum filtering and staged training
• Hard-question mining (e.g., pass_rate == 0.0)
• RL reward shaping / sampling strategies
• Controlled difficulty splits

---

Dataset filtering

This released version is filtered to remove trivial questions:

• All items with pass_rate == 1.0 are excluded

This keeps the dataset focused on non-trivial items with learning signal for RLVR / GRPO training.

---

Dataset structure

Columns

Field	Description
bolum	High-level subject / section (from turkish_mmlu)
konu	Topic under the subject (from turkish_mmlu)
soru	The question text (Turkish)
secenekler	List of answer options (MCQ choices)
cevap	Ground-truth correct option index (integer)
aciklama	Optional explanation / metadata (may be null)
pass_rate	Weak-model pass rate over parsable attempts (0.0–<1.0)

---

Intended use

Designed for:

• RLVR / GRPO training on Turkish MCQ reasoning
• Curriculum learning by difficulty (sampling via pass_rate)
• Hard-question filtering (e.g., pass_rate == 0.0)
• Turkish reasoning model evaluation on exam-style questions
• Weak-to-strong generalization experiments

Not intended for:

• Open-ended generation tasks without verification
• General factual QA outside the dataset’s question distribution
• Multimodal tasks

---

Citation

If you use this dataset, please credit the original source dataset:

• alibayram/turkish_mmlu

And mention the difficulty signal construction:

• Weak baseline: Qwen3-4B (no-thinking)
• 4 samples per question
• attempted counts only parsable outputs
• pass_rate = correct / attempted (0 if attempted=0)
• pass_rate == 1.0 filtered out in this release