kadirnar/kejue-slm-eval

Kejue SLM Eval — Kejue Bench v0.1

Kejue's standing company benchmark for the speech SLM. Private, frozen, versioned. Nothing in this dataset may enter any fine-tuning run, or it stops measuring regression.

Phase A (this version): extraction branches — ASR, dialect accuracy, translation, audio reasoning — across English, MSA and Arabic dialects, plus reserved Tier-3 stubs. Multi-turn and tool-calling (Phase B) come later.

Per-cell counts and confidence

Confidence flags are inline, next to the counts they qualify.

Cell	Source(s)	Split used	Raw avail	ASR	ASR target	Dialect	Dialect target	Confidence
english	LibriSpeech test-clean+test-other (openslr/librispeech_asr)	test-clean + test-other (held out of training)	5559	500	500	—	—	✅ full
msa	FLEURS ar_eg test+validation (google/fleurs)	test + validation	723	423	500	300	300	⚠️ under target — FLEURS ar_eg ceiling (723 clips total). Number is representative; smaller n → wider confidence band.
egyptian	Casablanca EGY (UBC-NLP/Casablanca:Egypt)	test + validation	1692	1000	1000	400	400	✅ full
levantine	Casablanca Jordan+Palestine, Syrian podcast (oddadmix/...-syrian-podcast)	test + validation (Casablanca) / train (Syrian podcast)	29349	1000	1000	400	400	✅ full (Jordanian+Palestinian+Syrian, balanced)
emirati	Casablanca UAE, Mixat (sqrk/mixat-tri)	test + validation (Casablanca) / test + train (Mixat)	6938	1000	1000	400	400	✅ full (Casablanca UAE + Mixat)
maghrebi	Casablanca Morocco+Algeria+Mauritania, Algerian podcasts (oddadmix/...loubna-stories/kahwa-postcast/rawi)	test + validation (Casablanca) / train (podcasts)	82833	1000	1000	400	400	✅ full (Moroccan+Algerian+Mauritanian+Algerian-podcast, balanced)
yemeni	Casablanca YEM (UBC-NLP/Casablanca:Yemen)	test + validation	1606	900	900	300	300	✅ full
saudi	HuggingPanda/Saudi_Podcasts_ASR	train	1009	445	800	300	300	⚠️ under target — Saudi Podcasts ceiling (MASC Saudi unavailable: dialect labels only in a 169GB login-gated tarball). Representative; smaller n → wider confidence band.
sudanese	AymanMansour/SDN-Dialect-Dataset-Filtered	train	3371	250	250	150	150	⚠️ LOW SPEAKER DIVERSITY — few distinct voices seen repeatedly. Less representative of Sudanese broadly than the 250 count suggests; speaker diversity (not clip count) is the thing to improve later.
libyan	oddadmix/arabic-audio-collection-libyan-a7rar-podcast	train	5357	1000	1000	150	150	⚠️ LOW SPEAKER DIVERSITY — single podcast series, few voices seen many times. Less representative of Libyan broadly than the 1000 count suggests; improve speaker diversity (not clip count) later.
translation	CoVoST2 ar->en TEST split (fixie-ai/covost2)	test (train split IS in training — excluded)	1695	1000	1000	—	—	✅ full
audio_reasoning	Big Bench Audio (fixie-ai/big_bench_audio)	test	1000	999	1000	—	—	✅ full (999/1000; 1 clip >60s cap)

Totals: 9517 ASR/translation/reasoning items + 2800 dialect-accuracy items = 12317 total.

Sources and licences

Cell	Source	Licence	Audio origin
english	LibriSpeech test-clean+test-other (openslr/librispeech_asr)	CC-BY-4.0	read audiobooks (held-out test splits, not in training)
msa	FLEURS ar_eg test+validation (google/fleurs)	CC-BY-4.0	read speech (not in training; CommonVoice-ar excluded as it IS trained)
egyptian	Casablanca EGY (UBC-NLP/Casablanca:Egypt)	CC-BY-NC-ND-4.0	YouTube TV-series
levantine	Casablanca Jordan+Palestine, Syrian podcast (oddadmix/...-syrian-podcast)	mixed (NC / other)	YouTube TV + human podcast (halabi2016 dropped: script-loaded)
emirati	Casablanca UAE, Mixat (sqrk/mixat-tri)	CC-BY-NC(-ND/-SA)	YouTube TV + code-switch podcast
maghrebi	Casablanca Morocco+Algeria+Mauritania, Algerian podcasts (oddadmix/...loubna-stories/kahwa-postcast/rawi)	mixed	YouTube TV + human podcasts
yemeni	Casablanca YEM (UBC-NLP/Casablanca:Yemen)	CC-BY-NC-ND-4.0	YouTube TV-series (cleanest cell: ar-ye not in training)
saudi	HuggingPanda/Saudi_Podcasts_ASR	unspecified	Saudi interview podcast (MASC Saudi top-up not possible: its dialect labels ship only in a 169GB login-gated tarball, stripped from the HF mirror)
sudanese	AymanMansour/SDN-Dialect-Dataset-Filtered	unspecified	human-recorded SDN corpus (synthetic Lisan dropped)
libyan	oddadmix/arabic-audio-collection-libyan-a7rar-podcast	other	human Libyan podcast (single series → speaker-limited)
translation	CoVoST2 ar->en TEST split (fixie-ai/covost2)	CC-BY-NC-4.0	read speech (held-out test split; train split IS in training)
audio_reasoning	Big Bench Audio (fixie-ai/big_bench_audio)	MIT	spoken reasoning questions (not in training)

Contamination controls

• Source blocklist (hard-fail): every clip is tagged with source_dataset; any source matching ADI-17, MGB-2/3/5, SADA22/sawtarabi, OmniLingual, MohamedRashad code-switching, QASR, any KejueAI/*, or NADI subtask-1/3 aborts the build. No resolved source hit it.
• Split discipline: LibriSpeech uses test splits only (training used train.*); CoVoST2 uses the ar_en test split only (training used ar_en train); FLEURS is absent from training. CommonVoice-ar was excluded — it IS in training.
• video_id dedup: podcast/YouTube cells (the oddadmix podcasts) are deduped against 4,332 real training YouTube video_ids extracted from KejueAI/arabic-youtube-corpus-with-continuations + KejueAI/lebanese-youtube-corpus.
• Disjoint ASR/dialect audio: within every cell the ASR and dialect-accuracy clips are drawn from a shared pool then partitioned — no audio is reused across branches (verified 0 overlap).
• Yemeni is the cleanest Arabic cell (ar-ye not in training). MASC is NOT in training (verified) so it would have been a legal Saudi source, but its dialect labels are unavailable (see Saudi note).

Normalisation and audio

• Text (verbatim from training): normalize_arabic_text then a has_word_repeated_3x row-drop on diacritic-intact text — both lifted byte-for-byte from the training pipeline with an import-time drift guard. strip_overlap is not ported (no continuation field); quotes wash out at scoring.
• Scoring normalisation (frozen, v1): stored per row as metadata.transcript_wer_norm — strips diacritics/hamza/madda, canonicalises alef+ya, Eastern→Western digits, strips punctuation except % and @; applied symmetrically to reference and hypothesis at scoring time (WER + CER).
• Audio: 16 kHz mono 16-bit PCM; 60s length cap (clips over 60s dropped — we cannot split a gold transcript to a sub-segment); music/heavy-noise/silence/clipping heuristics applied. Every dropped clip is logged with a reason.

Tier-3 stubs (0 rows, structure reserved)

These have no clean, non-contaminated open data yet and are not filled (no padding, no TTS, no substitution):

• kuwaiti — Tier 3: no clean open source; commercial (Nexdata/Pangeanic). Awaiting purchase.
• omani — Tier 3: commercial only. NOT OmniLingual (contaminated/in training). Awaiting purchase.
• bahraini — Tier 3: own source already in training (KejueAI/bahraini-speech). Need provably different-origin.
• iraqi — Deferred to v1.1: Appen/LDC2006S45 commercial telephone speech only.
• qatari — Deferred to v1.1: ArabicNLP-2025 corpus unsegmented, no timestamps, no pullable audio.

Phase B — Multi-turn (authored)

Each item is a real dialect/English audio clip used as the graded final user turn, preceded by an authored text conversation history that leads into it. The reference is a rubric (never a model-authored gold answer); scored later by the Gemini judge (scoring_method=judge_rubric).

• Author model: google/gemma-4-31b-it-20260402 (Gemma 4 31B via OpenRouter). Deliberately NOT Gemini — author and judge stay separate.
• Validator model (QC gate): qwen/qwen-2.5-72b-instruct (Qwen 2.5 72B) — a third, independent model (neither author nor judge) scored every item.
• Disjointness: multi-turn clips are disjoint (exact wav-hash + transcript) from the v0.1 asr/dialect clips in every cell except Saudi (see flag below). No new source corpora — same v0.1 sources, plus the untapped FLEURS-train split for MSA.

Cell	Multi-turn items	Flag
english	100	✅
msa	100	ℹ️ STAGED — FLEURS-train encyclopedic reads; histories frame the clip as an asserted fact the assistant reacts to (different flavour from the conversational dialect cells). Only clean disjoint MSA source available.
egyptian	100	✅
levantine	100	✅
emirati	100	✅
saudi	100	⚠️ SHARED AUDIO — reuses v0.1 Saudi clips as graded turns (no clean disjoint/topup Saudi source exists). Saudi MT and Saudi ASR/dialect scores are therefore NOT fully independent.
yemeni	100	✅
sudanese	100	✅
maghrebi	100	✅
libyan	100	✅

Total multi-turn: 1000 items across 10 languages.

Authoring QC pipeline (pre/post) and drop report

PRE (source-clip selection): 2s duration floor, 60s cap, audio/text QC, question/statement preference, dialect-label sanity, disjointness (Saudi excepted). POST (every authored item): deterministic checks (structural, clip-leak, encoding/mojibake, rubric shape, near-duplicate Jaccard≥0.85) then a Qwen validator call (coherence ≥3/5, dialect register, clip-contradiction, rubric gradeability). A circuit breaker halted any cell whose SYSTEMIC quality-drop rate exceeded 5% (with a minimum-evidence floor so 1-2 noise drops don't trip a full cell) for manual review.

Quality-drop rate: 0.004 (4/1004 attempts), 0 infra errors. Drops by reason: {'coherence': 3, 'register': 1}. The original 63 review items were re-validated through the same pipeline (3 defective ones dropped and backfilled), so all items meet one standard.

Phase B — Tool calling (audio commands → gold tool calls)

Spoken commands paired with a shared tool set and hand-authored gold calls. Scored programmatically (exact_tool_match), no judge. Taxonomy (VoiceAgentBench): single, retrieval (near-duplicate discrimination), parallel, sequential-dependent, and no-call.

Per-item fields for scoring: every row carries a top-level category column ∈ {single, retrieval, parallel, sequential, no_call} (queryable, consistent across English and all Arabic splits) so scores can be broken down by category, not just blended. The gold target is reference.expected_calls and the grading basis is the scoring_method column (exact_tool_match) — separate from the category label. The finer no-call sub-type is kept in metadata.call_type (safety_no_call for the English VoiceAgentBench safety items vs. none for benign Arabic no-call).

• English: extracted directly from krutrim-ai-labs/VoiceAgentBench (its real recorded English audio + its gold calls, preserved verbatim — arg values are lists of acceptable values; sequential items nest dependent calls). Its safety split is included as safety_no_call (no tool must fire). Passed the standard contamination gate.
• Arabic (all dialects): OUR OWN generated items, VoiceAgentBench-STYLE but NOT copied/translated — fresh scenarios per dialect from a validated generator (15 tools, varied arg types, near-duplicate tools; gold calls hand-constructed, 0 schema-validation errors). Gemma (google/gemma-4-31b-it-20260402) phrases the spoken dialectal command; Qwen (qwen/qwen-2.5-72b-instruct) validates faithfulness (all gold-arg facts expressed) + dialect register per item.
• Arabic TTS: google/gemini-3.1-flash-tts-preview (voice Kore, accent-as-prompt) for ALL dialects — chosen by a native-ear bake-off against x-ai/grok-voice-tts-1.0 and microsoft/mai-voice-2 on 9 varieties (bake-off assets kept in the repo tree).

Cell	Items	Audio	Flag
english	565	real recorded (VoiceAgentBench)	✅
msa	96	Gemini TTS	✅ TTS accent ear-confirmed (bake-off)
egyptian	134	Gemini TTS	✅ TTS accent ear-confirmed (bake-off)
levantine	135	Gemini TTS	✅ TTS accent ear-confirmed (bake-off)
emirati	135	Gemini TTS	✅ TTS accent ear-confirmed (bake-off)
saudi	135	Gemini TTS	✅ TTS accent ear-confirmed (bake-off)
yemeni	135	Gemini TTS	✅ TTS accent ear-confirmed (bake-off)
sudanese	134	Gemini TTS	✅ TTS accent ear-confirmed (bake-off)
maghrebi	133	Gemini TTS	✅ TTS accent ear-confirmed (bake-off)
libyan	135	Gemini TTS	✅ TTS accent ear-confirmed (bake-off)
kuwaiti	135	Gemini TTS	⚠️ SYNTHETIC AUDIO, dialect UNVERIFIED against native reference (no source audio exists for this dialect; TTS accent is a best-effort attempt). Tool-calling only — the ASR/dialect cells for this dialect remain empty stubs.
omani	135	Gemini TTS	⚠️ SYNTHETIC AUDIO, dialect UNVERIFIED against native reference (no source audio exists for this dialect; TTS accent is a best-effort attempt). Tool-calling only — the ASR/dialect cells for this dialect remain empty stubs.
bahraini	135	Gemini TTS	⚠️ SYNTHETIC AUDIO, dialect UNVERIFIED against native reference (no source audio exists for this dialect; TTS accent is a best-effort attempt). Tool-calling only — the ASR/dialect cells for this dialect remain empty stubs.
qatari	134	Gemini TTS	⚠️ SYNTHETIC AUDIO, dialect UNVERIFIED against native reference (no source audio exists for this dialect; TTS accent is a best-effort attempt). Tool-calling only — the ASR/dialect cells for this dialect remain empty stubs.
iraqi	133	Gemini TTS	⚠️ SYNTHETIC AUDIO, dialect UNVERIFIED against native reference (no source audio exists for this dialect; TTS accent is a best-effort attempt). Tool-calling only — the ASR/dialect cells for this dialect remain empty stubs.

Total tool_call: 2409 items across 15 splits.

Multi-turn tool calls (category multiturn_tool)

A reasoning-over-history category (20/dialect × 13 Arabic dialects): the text history contains prior tool calls across several turns, and the final AUDIO turn requires inferring the correct action from that history (e.g. "check the last category we haven't checked yet" → the model must scan the history and call get_offers with the one un-checked enum value). Final action is one / two / no call. Still programmatically scored on the gold call — NOT judge-scored. category="multiturn_tool", metadata.subset="multiturn_tool", metadata.scenario_type records the template.

Audio diversification (metadata.subset="audio_diversified")

15 extra rows/dialect × 13, evenly spread across the five categories (3 each), where ONLY the audio is diversified — the gold calls are normal. Each row pairs an emotion (surprised/angry/happy/sad/crying/…) with an acoustic condition (clean / white-noise / static-hiss / background-chatter / gain-low / gain-high), logged in metadata.emotion and metadata.audio_condition so tool-selection accuracy can be broken down by condition. Tests robustness of tool selection through emotion and noise.

Tier-3 note (honesty guard): generating tool-calling audio for kuwaiti/omani/bahraini/qatari/iraqi does NOT mean the benchmark now has ASR or dialect-accuracy data for them — those cells remain empty stubs. Tool calling scores tool selection, not accent fidelity, which is why synthetic carrier audio is acceptable here and nowhere else.

Audio reasoning — Arabic (translated Big Bench Audio)

Dialectal-Arabic reasoning items built with the same methodology as the Arabic tool-calling branch (no audio diversification): a random sample of Big Bench Audio questions is translated into each dialect by Gemma (preserving every premise, name, and number — the reasoning answer is unchanged), validated by Qwen for faithfulness + dialect register, then rendered to audio with Gemini TTS (google/gemini-3.1-flash-tts-preview, accent-as-prompt). Target 300/dialect.

• Gold answer is the original English gold_answer (translating the question does not change the answer). Scoring is identical to the English split: model_graded, Gemini judge. metadata carries bba_id, category, source_lang=en, translated=true.
• Same 4 Big Bench Audio categories (formal_fallacies, navigate, object_counting, web_of_lies), randomly mixed per dialect.

Dialect	Items
egyptian	300
levantine	300
emirati	300
saudi	300
yemeni	300
sudanese	300
maghrebi	300
libyan	300
kuwaiti	300
omani	300
bahraini	300
qatari	300
iraqi	300

Arabic audio-reasoning total so far: 3900 across 13 dialects (English split unchanged at 999). Remaining dialects land additively as they finish.

Schema (one unified schema, all branches)

id, task, language, region, source_dataset, audio (16kHz), audio_transcript, system_prompt, history, tools, reference, scoring_method, judge_model, metadata. Configs: asr, dialect, translation, audio_reasoning, multi_turn, tool_call, split by language.

Version

Tag v0.1. Frozen. Phase-B branches (multi_turn and, later, tool_call) are additive configs layered on the frozen v0.1 extraction cells — no v0.1 row is edited. Tier-3 dialects and any Saudi/MSA top-ups likewise land additively.