CLUAS: Comprehensive Listening and Understanding ASsessment is an Irish language Speech LLM evaluation dataset based on the Higher Level Irish Leaving Certificate listening-comprehension papers for 2013–2025.

13 exam years and 780 available marks (60 per year)
104 , 16 kHz mono WAV audio clips (context)
334 questions (text) grouped by their respective audio clip
gold transcript
marking-scheme rubrics

Exam Structure

The structure and mark allocation are consistent across all 13 exam years:

Section	Material	Dataset clips	Marks
Cuid A	Two announcements (`Fógra 1` and `Fógra 2`)	2	18
Cuid B	Two conversations (`Comhrá 1` and `Comhrá 2`), each divided into two `mír` question clips	4	28
Cuid C	Two spoken pieces (`Píosa 1` and `Píosa 2`)	2	14
Total	Six exam items represented by eight question-bearing clips	8	60

Announcements and pieces are heard twice in the exam. Each conversation is first heard in full and is then heard in two parts. CLUAS uses the question-bearing hearing for each clip. The number of question records varies between years because an exam question may contain several separately marked parts, but the section and overall mark totals do not change.

The official schemes award 60 content marks (Eolas) and then allow a separate overall deduction of 0–3 marks for the standard of Irish (Cumas Gaeilge). CLUAS currently stores and reports the 60 content marks question by question; it does not model that separate holistic language deduction.

Loading

from datasets import load_dataset

ds = load_dataset("jmcinern/CLUAS", split="test")
example = ds[0]
audio = example["audio"]
for question in example["questions"]:
    print(question["question_id"], question["question"])

The file_name field links the primary clip. context_audio_file_name is null for announcements and pieces; for each conversation half it links the earlier full hearing of that conversation.

Evaluation protocol

Run the same model in three conditions on every question:

Baseline: no_context: the models baseline performance with no context. Some of the questions are quite guessable / refer to real world events.
just_audio: the models performance with the speech clip as context
just_transcript: the models performance with the marking scheme provided transcript

Evaluate each free-text answer with a separate LLM judge. Give the judge the question, candidate answer, marking_scheme, marks_available, and blanks_required. Never give the marking scheme or gold transcript to the model being evaluated.

The functions answer_with_llm and judge_with_llm below stand for calls to your preferred speech LLM and judge APIs. Configure them with your API key and model.

import json
from datasets import load_dataset

ds = load_dataset("jmcinern/CLUAS", split="test")

JUDGE_PROMPT = """You are an experienced examiner marking An Chluastuiscint,
the listening-comprehension section of the Irish Leaving Certificate Higher Level
Irish examination.

Apply these rules strictly:
1. The candidate must answer in Irish. If the substantive answer is in English,
   award 0. Do not reject language-neutral content such as a name, place, date,
   number, currency, URL, or email address merely because it is not an Irish word.
2. Mark the meaning expressed in Irish, not exact string overlap. Accept equivalent
   Irish wording, dialect forms, inflections, and minor spelling or grammar errors
   when the intended answer is unambiguous. Do not apply a separate language-quality
   deduction: this dataset reports the scheme's 60 content marks only.
3. Follow only the supplied marking scheme. A tier marked '= 2 mharc' earns two
   marks, '= 1 mharc' earns one, and '= 0' is explicitly refused. Do not upgrade a
   partial answer because it sounds plausible or appears in the transcript.
4. Several same-value alternatives normally mean that any one earns that tier.
   Parenthetical text may be optional qualification or an examiner instruction;
   interpret it in context.
5. For questions requiring multiple facts, credit only distinct correct points.
   Rephrasing the same fact twice fills one blank. Never exceed blanks_required or
   marks_available.
6. Obey explicit scheme notes, including required units/details and cancellation
   rules where a correct and contradictory answer are both supplied. Ignore stray
   PDF page numbers or broken line wrapping.
7. Do not award marks for information absent from the marking scheme, even if it
   is true of the recording.

Return JSON only:
{"marks": <integer>, "matched_tiers": ["<verbatim credited scheme tier>"],
 "reason": "<brief explanation in English>"}."""

results = []
for example in ds:
    for question in example["questions"]:
        for condition in ("no_context", "just_audio", "just_transcript"):
            answer = answer_with_llm(
                question=question["question"],
                audio_path=example["file_name"] if condition == "just_audio" else None,
                transcript=example["gold_transcript"] if condition == "just_transcript" else None,
            )

            verdict = judge_with_llm(
                system_prompt=JUDGE_PROMPT,
                payload={
                    "question": question["question"],
                    "candidate_answer": answer,
                    "marking_scheme": question["marking_scheme"],
                    "marks_available": question["marks_available"],
                    "blanks_required": question["blanks_required"],
                },
            )
            results.append({
                "question_id": question["question_id"],
                "condition": condition,
                "answer": answer,
                "marks": verdict["marks"],
            })

with open("results.json", "w", encoding="utf-8") as f:
    json.dump(results, f, ensure_ascii=False, indent=2)

Example

Question ID	Question number	Question	Marking scheme	Marks
`2013-B-comhra1-q1`	1	Cad a rinne Orla inniu?	Chríochnaigh sí an Ardteist = 2 mharc; Bhí an scrúdú deireanach aici inniu = 2 mharc; Bitheolaíocht = 2 mharc; Scrúdú san Ardteist = 1 mharc; Ardteist = 0	2