Datasets:
Add gpt-oss-120B + REAP; withhold pre-fix HumanEval (reasoning-aware harness)
Browse files
README.md
CHANGED
|
@@ -33,14 +33,22 @@ Speed and quality benchmarks for quantized LLMs on NVIDIA RTX 5090 32GB, measure
|
|
| 33 |
|
| 34 |
Generative evaluation through llama-server chat completions. Replicates standard benchmark methodology using custom evaluators — no `lm-evaluation-harness` dependency.
|
| 35 |
|
| 36 |
-
| Model | Params | Quant | MMLU | ARC-C | HellaSwag | GSM8K | HumanEval |
|
| 37 |
|-------|-------:|-------|-----:|------:|----------:|------:|----------:|
|
| 38 |
-
|
|
| 39 |
-
| Qwen3.6-
|
| 40 |
-
|
|
| 41 |
-
| Qwen3-
|
| 42 |
-
|
|
| 43 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 44 |
|
| 45 |
> All models benchmarked with 50% stratified sampling (seed=42), thinking disabled. Full per-model reports in [`reports/`](reports/).
|
| 46 |
|
|
|
|
| 33 |
|
| 34 |
Generative evaluation through llama-server chat completions. Replicates standard benchmark methodology using custom evaluators — no `lm-evaluation-harness` dependency.
|
| 35 |
|
| 36 |
+
| Model | Params | Quant | MMLU | ARC-C | HellaSwag | GSM8K | HumanEval¹ |
|
| 37 |
|-------|-------:|-------|-----:|------:|----------:|------:|----------:|
|
| 38 |
+
| gpt-oss-120B² | 116.83B | MXFP4 | 89.5 | 95.0 | 80.0 | 97.0 | 98.0 |
|
| 39 |
+
| Qwen3.6-28B-REAP-A3B | 28.24B | Q6_K | 87.7 | 95.0 | 82.0 | 90.0 | 94.0 |
|
| 40 |
+
| Gemma 4 31B-it | 30.70B | Q6_K | 87.8 | 97.6 | 92.0 | 97.5 | — |
|
| 41 |
+
| Qwen3.6-27B | 26.90B | Q6_K | 87.9 | 96.9 | 95.4 | 97.3 | — |
|
| 42 |
+
| Qwen3.6-35B-A3B | 34.66B | UD-Q4_K_M | 85.0 | 95.7 | 93.4 | 96.7 | — |
|
| 43 |
+
| Qwen3-Coder-Next | 79.67B | UD-Q2_K_XL | 83.7 | 96.0 | 89.3 | 96.0 | — |
|
| 44 |
+
| Nemotron-Cascade-2 | 31.58B | Q4_K_M | 74.4 | 91.5 | 75.7 | 87.1 | — |
|
| 45 |
+
| gpt-oss-20b | 20.91B | Q4_K_M | 78.6 | 94.6 | 74.5 | 94.8 | — |
|
| 46 |
+
|
| 47 |
+
> ¹ **HumanEval** is shown only for models run on the **corrected, reasoning-aware harness** (no premature stop sequences, indentation-preserving response handling). The earlier harness systematically **understated** models that reason inline, so those rows' HumanEval is **withheld pending re-run** — do not cite the prior figures. gpt-oss-120B and Qwen3.6-28B-REAP use the fixed harness.
|
| 48 |
+
>
|
| 49 |
+
> ² **gpt-oss-120B** runs via MoE CPU-offload (`--n-cpu-moe 20`) — it does not fit 32GB VRAM (59GB model); ~30GB VRAM + the rest in system RAM, ~47 tok/s generation vs the full-VRAM models.
|
| 50 |
+
>
|
| 51 |
+
> The two top rows were run **reasoning-on** on a ~100-item-per-task subset (MMLU 2/subject); earlier rows used the full item counts in the Methodology table, reasoning-off. The two sets are **not directly comparable** on MCQ/GSM8K either.
|
| 52 |
|
| 53 |
> All models benchmarked with 50% stratified sampling (seed=42), thinking disabled. Full per-model reports in [`reports/`](reports/).
|
| 54 |
|