Datasets:

johnbean393
/

fluid-2-sft-eval

@@ -96,23 +96,6 @@ 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.
-## Reference 0.8B beta result
-The selected
-[`johnbean393/fluid-2-qwen3.5-0.8b-beta`](https://huggingface.co/johnbean393/fluid-2-qwen3.5-0.8b-beta)
-checkpoint produced these corrected deterministic greedy scores:
-| System | Scored rows | Exact match | CER | WER |
-|---|---:|---:|---:|---:|
-| Fluid 2 Qwen3.5 0.8B Beta | 7,010 | **27.6605%** | **17.8294%** | **28.2930%** |
-| Raw `asr_text` copy baseline | 7,010 | 3.9372% | 39.0639% | 60.0924% |
-The model reduces CER by 54.36% and WER by 52.92% relative to the raw-ASR copy
-baseline on the identical scoring cohort. The primary denominator excludes 121
-EOS-only rows and 30 generations that reached the 4,096-token completion limit.
-Both cohorts remain in the full 7,161-request evaluation and are reported as
-separate termination diagnostics.
 The corpus combines sources with different upstream terms. The parent dataset
 card documents provenance and license labels; this repository does not replace
 those terms with a blanket license.
@@ -120,15 +103,15 @@ those terms with a blanket license.
 <!-- FLUID2_EVAL_RESULTS_START -->
 ## Published model results
-All results below use deterministic greedy SGLang inference. EM, CER, and WER 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 | WER | Evaluated revision |
-|---|---:|---:|---:|---:|---:|---:|---|
-| [0.8B Beta](https://huggingface.co/johnbean393/fluid-2-qwen3.5-0.8b-beta) | 7,010 | 121 | 30 | 27.6605% | 17.8294% | 28.2930% | `e42c24cc3b71` |
-| [2B Beta](https://huggingface.co/johnbean393/fluid-2-qwen3.5-2b-beta) | 7,016 | 121 | 24 | 30.3449% | 17.4199% | 27.4544% | `f71a0445a8f1` |
-| [4B Beta](https://huggingface.co/johnbean393/fluid-2-qwen3.5-4b-beta) | 7,024 | 121 | 16 | 33.5849% | 16.4947% | 25.4750% | `f6e0aaa0dac4` |
-| [0.8B Beta1](https://huggingface.co/johnbean393/fluid-2-qwen3.5-0.8b-beta1) | 7,011 | 121 | 29 | 27.5567% | 18.7717% | 29.1997% | `0e152e2591c7` |
-| [2B Beta1](https://huggingface.co/johnbean393/fluid-2-qwen3.5-2b-beta1) | 7,020 | 121 | 20 | 28.8177% | 18.7955% | 28.9881% | `d433d5f31a24` |
-| [4B Beta1](https://huggingface.co/johnbean393/fluid-2-qwen3.5-4b-beta1) | 7,030 | 121 | 10 | 32.0341% | 17.5259% | 27.1583% | `a02b1fcb9972` |
 <!-- FLUID2_EVAL_RESULTS_END -->

 text-quality scores. Throughput and inference-failure counts still cover the full
 7,161-request census.
 The corpus combines sources with different upstream terms. The parent dataset
 card documents provenance and license labels; this repository does not replace
 those terms with a blanket license.
 <!-- FLUID2_EVAL_RESULTS_START -->
 ## Published model results
+All results below use deterministic greedy SGLang inference. 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 |
+|---|---:|---:|---:|---:|---:|---|
+| [0.8B Beta](https://huggingface.co/johnbean393/fluid-2-qwen3.5-0.8b-beta) | 7,010 | 121 | 30 | 27.6605% | 17.8294% | `e42c24cc3b71` |
+| [2B Beta](https://huggingface.co/johnbean393/fluid-2-qwen3.5-2b-beta) | 7,016 | 121 | 24 | 30.3449% | 17.4199% | `f71a0445a8f1` |
+| [4B Beta](https://huggingface.co/johnbean393/fluid-2-qwen3.5-4b-beta) | 7,024 | 121 | 16 | 33.5849% | 16.4947% | `f6e0aaa0dac4` |
+| Raw `asr_text` copy baseline (0.8B cohort) | 7,010 | 121 | 30* | 3.9372% | 39.0639% | — |
+\* 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.
 <!-- FLUID2_EVAL_RESULTS_END -->