JimHue/TASTE-S-EXP-Result

TASTE-S QA Evaluation Results

This repository contains selected QA evaluation artifacts for pretrained LLM baselines and turn-based TASTE-S SLMs. Each entry is one model × input condition. S2T scores text answers; S2S scores answers transcribed from generated speech. Both use UltraEval qa-exist-match exact match. Dataset sizes are LLaMA Questions (300), TriviaQA (1,024), and WebQuestions (2,032).

Results

Model	Question/input condition	LLaMA Questions	TriviaQA	WebQuestions	Result directory
Qwen2.5-7B-Instruct	Original benchmark question text (UltraEval T2T)	71.00% (213/300)	54.10% (554/1,024)	49.26% (1,001/2,032)	Qwen2.5_7B_Instruct/evaluation_result/spokenqa/ultraeval_t2t
Qwen3.5-0.8B	Original question text, no instruction	47.00% (141/300)	Not evaluated	Not evaluated	Qwen3.5_0.8B/evaluation_result/spokenqa/qwen35_0_8b_text_only_no_instruction_v1
Qwen3.5-0.8B	Historical CTC transcript input v1	37.67% (113/300)	15.33% (157/1,024)	15.11% (307/2,032)	Qwen3.5_0.8B/evaluation_result/spokenqa/qwen35_0_8b_text_only_ctc_input_v1
Qwen3.5-0.8B	Whisper Large V3 transcript	47.67% (143/300)	18.26% (187/1,024)	25.20% (512/2,032)	Qwen3.5_0.8B/evaluation_result/spokenqa/qwen_whisper_large_v3_asr
Qwen3.5-0.8B	New tokenizer CTC transcript (EXP-TOK-039)	48.00% (144/300)	17.38% (178/1,024)	23.67% (481/2,032)	Qwen3.5_0.8B/evaluation_result/spokenqa/qwen35_0_8b_new_tokenizer_ctc_EXP-TOK-039
Qwen3.5-4B	Original question text, no instruction	70.67% (212/300)	48.93% (501/1,024)	42.67% (867/2,032)	Qwen3.5_4B/evaluation_result/spokenqa/qwen35_4b_text_only_no_instruction_v1
Qwen3.5-4B	Historical CTC transcript input v1	57.67% (173/300)	32.42% (332/1,024)	27.02% (549/2,032)	Qwen3.5_4B/evaluation_result/spokenqa/qwen35_4b_text_only_ctc_input_v1
Qwen3.5-4B	Whisper Large V3 transcript	70.67% (212/300)	41.31% (423/1,024)	41.49% (843/2,032)	Qwen3.5_4B/evaluation_result/spokenqa/qwen_whisper_large_v3_asr
Qwen3.5-4B	New tokenizer CTC transcript (EXP-TOK-039)	70.00% (210/300)	36.13% (370/1,024)	38.04% (773/2,032)	Qwen3.5_4B/evaluation_result/spokenqa/qwen35_4b_new_tokenizer_ctc_EXP-TOK-039
Qwen3.5-9B	Original question text, no instruction	74.33% (223/300)	56.93% (583/1,024)	48.97% (995/2,032)	Qwen3.5_9B/evaluation_result/spokenqa/qwen35_9b_text_only_no_instruction_v1
Qwen3.5-9B	Historical CTC transcript input v1	62.00% (186/300)	38.77% (397/1,024)	Not evaluated	Qwen3.5_9B/evaluation_result/spokenqa/qwen35_9b_text_only_ctc_input_v1
Qwen3.5-9B	Whisper Large V3 transcript	74.33% (223/300)	49.22% (504/1,024)	46.95% (954/2,032)	Qwen3.5_9B/evaluation_result/spokenqa/qwen_whisper_large_v3_asr
Qwen3.5-9B	New tokenizer CTC transcript (EXP-TOK-039)	74.33% (223/300)	44.82% (459/1,024)	43.06% (875/2,032)	Qwen3.5_9B/evaluation_result/spokenqa/qwen35_9b_new_tokenizer_ctc_EXP-TOK-039
Qwen3.5-0.8B	Parakeet TDT 0.6B v2 ASR transcript	46.67% (140/300)	17.29% (177/1,024)	25.00% (508/2,032)	Qwen3.5_0.8B/evaluation_result/spokenqa/qwen_parakeet_tdt_0_6b_v2_asr
Qwen3.5-4B	Parakeet TDT 0.6B v2 ASR transcript	69.67% (209/300)	38.87% (398/1,024)	40.55% (824/2,032)	Qwen3.5_4B/evaluation_result/spokenqa/qwen_parakeet_tdt_0_6b_v2_asr
Qwen3.5-9B	Parakeet TDT 0.6B v2 ASR transcript	74.33% (223/300)	47.36% (485/1,024)	45.37% (922/2,032)	Qwen3.5_9B/evaluation_result/spokenqa/qwen_parakeet_tdt_0_6b_v2_asr

Turn-based SLM results

Both turn-based evaluations below have full paired coverage on all three datasets (3,356 examples each in S2S and S2T). S2S is generated speech transcribed and scored; S2T is the model's text answer.

Model	Evaluation condition	LLaMA Questions (S2S / S2T)	TriviaQA (S2S / S2T)	WebQuestions (S2S / S2T)	Result directory
Qwen3.5-0.8B TASTE-S SLM (MODEL-SLM-024; EVAL-SLM-084)	Parakeet TDT v2 text + bidirectional FSQ	33.67% / 42.67%	13.18% / 15.72%	20.62% / 22.54%	Qwen3.5_0.8B/evaluation_result/spokenqa/turnbased_0p8b_fsdp_eob_w1_protected_spokenqa_all_parakeet_tdt_v2/aggregate
Qwen3.5-0.8B Tok014 TASTE-S SLM (MODEL-SLM-028; EVAL-SLM-088)	Parakeet TDT v2 text + Tok014 bidirectional FSQ; non-streaming decoder	41.00% / 49.33%	15.14% / 17.09%	22.29% / 23.23%	Qwen3.5_0.8B/evaluation_result/spokenqa/turnbased_0p8b_fsdp_eob_w1_protected_spokenqa_all_parakeet_asrtext_tok014/aggregate

Directory contents

The top-level directories group artifacts by Qwen backbone size. Under Qwen3.5_0.8B/evaluation_result/spokenqa/, baseline conditions and turn-based training runs have separate folders; each turn-based aggregate/ folder contains its summary files and consolidated full-dataset per-sample S2S/S2T predictions, with shard-level outputs omitted. MANIFEST.csv records the relative path, byte size, and SHA-256 digest of every copied result file.

Input conditions:

• Original question text, no instruction: canonical dataset question supplied directly to the pretrained Qwen model, without an extra instruction wrapper.
• Original benchmark question text (UltraEval T2T): benchmark question text evaluated as a text-to-text task with Qwen2.5-7B-Instruct.
• Whisper Large V3 transcript: question transcript produced by Whisper Large V3, supplied to the pretrained Qwen model.
• Parakeet TDT 0.6B v2 transcript: Parakeet ASR transcript of each spoken benchmark question, supplied as text input to the original pretrained Qwen3.5 model. These are S2T exact-match results; no speech output is generated.
• New tokenizer CTC transcript (EXP-TOK-039): transcript from the new bidirectional CTC tokenizer's valid_best checkpoint, supplied to the pretrained Qwen model. This is a diagnostic evaluation; these QA datasets also participate in tokenizer validation/checkpoint selection, so it is not independent held-out evidence. The checkpoint SHA-256 recorded for this run is e9361806f2d521d204195a46f6102f94ec1916c5a51edab3812f6ad0ceb19701.
• Historical CTC transcript input v1: archived earlier CTC-input evaluation. The archived files have limited run provenance; this label does not identify it as EXP-TOK-037 or EXP-TOK-039.

“Not evaluated” means no result was recorded for that dataset and must not be interpreted as zero. Historical text_only_ctc_input_v1 results for Qwen3.5-9B cover LLaMA Questions and TriviaQA only. Qwen3.5-0.8B's no-instruction run covers LLaMA Questions only.

The listed pretrained text-only LLM baselines report S2T only. Both turn-based SLM evaluations report paired S2S/S2T scores with full scored coverage and zero inference failures.

Scope and exclusions

This repository contains selected evaluation result folders grouped under the Qwen2.5-7B-Instruct and Qwen3.5 0.8B/4B/9B size directories. It includes selected pretrained text-only LLM controls and the two completed 0.8B turn-based SLM evaluations listed above. It excludes every EXP-TOK-037 run and excludes Qwen3.5_4B/evaluation_result/spokenqa/qwen35_4b_text_baseline. No model checkpoints, audio datasets, benchmark source datasets, or external dataset caches are included.