johnbean393/fluid-2-sft-eval

Fluid 2 — dictation cleanup eval

This repository contains 7,161 text-only voice-dictation cleanup evaluation rows for like-for-like model comparison. No audio is included or fetched.

Split	Rows	Documents	Audio
eval	7,161	2,718	Not included

What this benchmark tests

The benchmark measures whether a model can turn noisy voice dictation into the intended written text without answering it or adding content. The rows cover:

• local ASR, spelling, capitalization, punctuation, and grammar repair;
• filler removal, abandoned starts, cancellations, and correction chains;
• spoken punctuation plus line, paragraph, and list formatting;
• short notes and questions through long, multi-paragraph passages; and
• dictation with no surrounding context, one-sided context, or both previous and following context fields.

The 121 empty-target rows represent fully cancelled dictation and test whether the model stops without emitting cleaned text.

Measured content distribution

The following counts are computed over all 7,161 rows. Target-length counts include the 121 empty targets.

Clean-target length	Rows	Share
Empty	121	1.69%
1–50 characters	3,198	44.66%
51–150 characters	2,601	36.32%
151–500 characters	1,123	15.68%
501+ characters	118	1.65%

Available context	Rows	Share
Neither side	2,069	28.89%
Previous only	2,665	37.22%
Following only	655	9.15%
Previous and following	1,772	24.75%

The median ASR input is 14 words (95th percentile: 53; maximum: 596). Among non-empty targets, the median is 11 words (95th percentile: 44; maximum: 427). Of the 7,040 non-empty targets, 6,764 (96.08%) differ exactly from their ASR input; 5,068 are shorter, 1,389 are longer, and 583 have the same character length while still potentially differing in content.

Formatting behaviors are materially represented and the categories overlap: 1,090 targets contain a line break, 562 contain a blank-line paragraph break, 203 contain a Markdown-style list item, and 423 end in a question mark.

Schema and loading

Only the seven columns consumed by text SFT and evaluation are present:

• row_id, job_id, document_id
• prev_text, post_text, asr_text, clean_target

from datasets import load_dataset

eval_rows = load_dataset(
    "johnbean393/fluid-2-sft-eval",
    split="eval",
)
assert len(eval_rows) == 7_161

No audio, sampling-rate, duration, or other waveform-bearing column is included, so ordinary loading cannot fetch or decode audio.

Prompt templates

The Fluid 2 Beta rows are rendered as a raw completion prompt; do not apply a chat template:

<|dictation_clean_v1|>
<|start_prev_text|>{prev_text}<|end_prev_text|>
<|start_post_text|>{post_text}<|end_post_text|>
<|start_asr_text|>{asr_text}<|end_asr_text|>
<|start_target_text|>

The model generates clean_target and terminates at <|end_target_text|>.

The instruction-tuned prompt-only Qwen and Gemma references, their Fluid fine-tunes, DeepSeek V4 Flash, and GPT-5.6 Luna instead use chat messages. Qwen, Gemma, and the Fluid fine-tunes use each GGUF's native template; DeepSeek uses its official chat-completions API; Luna uses the local Codex Responses API. All receive the same private system instruction and the row's asr_text verbatim as the user message. The private instruction is intentionally not reproduced in this public repository.

Metrics

The Fluid 2 Beta rows use SGLang raw-completion inference. Instruction-tuned Qwen/Gemma comparison rows use llama.cpp chat-completion inference with their native GGUF templates and reasoning disabled during generation. DeepSeek uses the official API with thinking disabled. Those paths use temperature-zero decoding. Luna uses the local Codex Responses API with reasoning effort set to none and proxy/model sampling defaults. The public scoreboard reports strict exact match and character error rate (CER), while the audit also tracks whitespace-normalized exact match, improvement over the raw-ASR copy baseline, empty/control-token failures, and context-bucket breakdowns. Strict metrics remain case-, punctuation-, and formatting-sensitive because those are core behaviors of the cleanup model.

Rows with an empty clean_target supervise only the terminal control token. The harness reports their empty-completion/stop behavior separately and excludes them from the EM and CER denominators. Generations that reach the configured completion-token limit are likewise reported separately and excluded from those text-quality scores. Throughput and inference-failure counts still cover the full 7,161-request census.

Published model results

Fluid 2 Beta results use deterministic greedy SGLang raw-completion inference. The prompt-only Qwen and Gemma references and their Fluid fine-tunes use deterministic greedy llama.cpp chat-completion inference with each GGUF's native chat template and reasoning disabled at generation. DeepSeek V4 Flash uses its official chat-completions API with thinking disabled. GPT-5.6 Luna uses the local Codex Responses API with reasoning effort set to none. All nine models receive the same 7,161-row text-only development set; every chat model receives the same fixed private system instruction and verbatim asr_text user message. EM and CER exclude both EOS-only rows and generation-capped rows; the exclusion counts are shown explicitly.

Model	Scored rows	EOS-only excluded	Capped excluded	EM	CER	Evaluated revision
Fluid 2 Qwen3.5 0.8B Beta	7,010	121	30	27.6605%	17.8294%	e42c24cc3b71
Fluid 2 Qwen3.5 2B Beta	7,016	121	24	30.3449%	17.4199%	f71a0445a8f1
Fluid 2 Qwen3.5 4B Beta	7,024	121	16	33.5849%	16.4947%	f6e0aaa0dac4
Qwen3.5-2B (prompt only)	6,996	121	44	4.4025%	47.9775%	15852e8c1636
Fluid-1 Mini	7,023	121	17	5.2684%	28.1837%	fc9e22028bf2
Gemma 4 E2B IT (prompt only)	7,034	121	6	5.7293%	42.3142%	3e22461f65e8
FluidIntelligence	7,023	121	17	5.8522%	27.5391%	0980ee1c2019
DeepSeek V4 Flash (prompt only)	7,033	121	7	6.3984%	46.1081%	a26a7955944d
GPT-5.6 Luna (prompt only)	6,984	121	57	6.2285%	32.8253%	gpt-5.6-luna
Raw asr_text copy baseline (0.8B cohort)	7,010	121	30*	3.9372%	39.0639%	—

The llama.cpp comparison runs use temperature=0, top_k=1, top_p=1, min_p=0, a fixed request seed, a 4,096-token completion limit, no speculative decoding, and --reasoning off --reasoning-budget 0.

The DeepSeek run uses deepseek-v4-flash, temperature=0, top_p=1, a 4,096-token completion limit, and the official thinking: {type: disabled} switch. Its evaluated revision is the API-returned system fingerprint.

The GPT-5.6 Luna run uses the local Codex Responses API, model gpt-5.6-luna, proxy/model sampling defaults with no sampling override, reasoning.effort=none, a 4,096-token output limit, and a 64-request client-side in-flight ceiling. At the requested early stop, the 57 unreturned proxy-timeout rows are treated as capped exclusions; one of those rows is also EOS-only, so the exclusion columns are not disjoint for this run.

* The raw-ASR copy baseline uses the identical 7,010-row scoring cohort as the 0.8B model. Its capped-excluded count is inherited from the 0.8B generation outcomes; the copy baseline itself does not generate or hit a token cap.

Paired prompt-only comparison

Each delta below is recomputed on the identical intersection of rows eligible for both models: empty-target rows and any row capped by either member of the pair are excluded. Positive EM gain and positive relative CER reduction indicate improvement by the Fluid fine-tune.

Reference → Fluid fine-tune	Common scored rows	Union capped excluded	Reference EM	Fine-tuned EM	EM gain	Reference CER	Fine-tuned CER	Relative CER reduction
Qwen/Qwen3.5-2B → altic-dev/fluid-1-mini	6,994	46	4.4038%	5.2902%	+0.8865 pp	47.9434%	28.1521%	+41.2806%
google/gemma-4-E2B-it → altic-dev/FluidIntelligence	7,022	18	5.7391%	5.8530%	+0.1139 pp	41.9034%	27.5406%	+34.2760%

Both comparisons are size- and family-matched: Qwen 3.5 2B and Fluid-1 Mini are evaluated at Q6_K, while Gemma 4 E2B and FluidIntelligence are evaluated at Q4_K_M.