| { |
| "chart": "quantization", |
| "metric": "top-1 agreement with the full-precision reference (%), plus file size (GB = bytes/1e9)", |
| "rows": [ |
| { |
| "tier": "16-bit", |
| "model": "v3", |
| "label": "Jev-Style 0.8B v3 \u00b7 GGUF F16", |
| "agree": 100.0, |
| "n": 240, |
| "agree_count": 240, |
| "size_gb": 1.516744128, |
| "size_note": "local export file (bytes / 1e9)", |
| "reference": "torch FP32", |
| "source": "runs/macjev/export_r2/main/parity/report.json; sizes: chart_data.json export.sizes" |
| }, |
| { |
| "tier": "16-bit", |
| "model": "v3", |
| "label": "Jev-Style 0.8B v3 \u00b7 MLX bf16", |
| "agree": 100.0, |
| "n": 240, |
| "agree_count": 240, |
| "size_gb": 1.504827355, |
| "size_note": "local export file (bytes / 1e9)", |
| "reference": "torch FP32", |
| "source": "runs/macjev/export_r2/main/parity/report.json; sizes: chart_data.json export.sizes" |
| }, |
| { |
| "tier": "16-bit", |
| "model": "v2", |
| "label": "Jev-Style 2B v2 \u00b7 GGUF BF16", |
| "agree": 99.6, |
| "n": 500, |
| "size_gb": 3.78, |
| "size_note": "as reported on card", |
| "reference": "CUDA merged BF16", |
| "source": "https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-v2-GGUF/raw/main/README.md" |
| }, |
| { |
| "tier": "16-bit", |
| "model": "v2", |
| "label": "Jev-Style 2B v2 \u00b7 MLX BF16", |
| "agree": 99.6, |
| "n": 500, |
| "size_gb": 3.76, |
| "size_note": "as reported on card", |
| "reference": "CUDA merged BF16", |
| "source": "https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-v2-GGUF/raw/main/README.md (table: 'Native MLX BF16 | 3.76 GB | 99.6% choice agreement')" |
| }, |
| { |
| "tier": "16-bit", |
| "model": "v1", |
| "label": "Jev-Style 2B v1 \u00b7 GGUF BF16", |
| "agree": 99.8, |
| "n": 500, |
| "size_gb": 3.9, |
| "size_note": "as reported on card", |
| "reference": "bf16", |
| "source": "https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-GGUF/raw/main/README.md" |
| }, |
| { |
| "tier": "8-bit", |
| "model": "v3", |
| "label": "Jev-Style 0.8B v3 \u00b7 GGUF Q8_0", |
| "agree": 100.0, |
| "n": 240, |
| "agree_count": 240, |
| "size_gb": 0.811843008, |
| "size_note": "local export file (bytes / 1e9)", |
| "reference": "torch FP32", |
| "source": "runs/macjev/export_r2/main/parity/report.json; sizes: chart_data.json export.sizes" |
| }, |
| { |
| "tier": "8-bit", |
| "model": "v3", |
| "label": "Jev-Style 0.8B v3 \u00b7 MLX 8-bit", |
| "agree": 100.0, |
| "n": 240, |
| "agree_count": 240, |
| "size_gb": 0.799973748, |
| "size_note": "local export file (bytes / 1e9)", |
| "reference": "torch FP32", |
| "source": "runs/macjev/export_r2/main/parity/report.json; sizes: chart_data.json export.sizes" |
| }, |
| { |
| "tier": "8-bit", |
| "model": "v2", |
| "label": "Jev-Style 2B v2 \u00b7 GGUF Q8_0", |
| "agree": 99.2, |
| "n": 500, |
| "size_gb": 2.01, |
| "size_note": "as reported on card", |
| "reference": "CUDA merged BF16", |
| "source": "https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-v2-GGUF/raw/main/README.md" |
| }, |
| { |
| "tier": "8-bit", |
| "model": "v1", |
| "label": "Jev-Style 2B v1 \u00b7 GGUF Q8_0", |
| "agree": 99.4, |
| "n": 500, |
| "size_gb": 2.1, |
| "size_note": "as reported on card", |
| "reference": "bf16", |
| "source": "https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-GGUF/raw/main/README.md" |
| }, |
| { |
| "tier": "4-bit", |
| "model": "v3", |
| "label": "Jev-Style 0.8B v3 \u00b7 GGUF Q4_K_M", |
| "agree": 100.0, |
| "n": 240, |
| "agree_count": 240, |
| "size_gb": 0.529296832, |
| "size_note": "local export file (bytes / 1e9)", |
| "reference": "torch FP32", |
| "source": "runs/macjev/export_r2/main/parity/report.json; sizes: chart_data.json export.sizes" |
| }, |
| { |
| "tier": "4-bit", |
| "model": "v2", |
| "label": "Jev-Style 2B v2 \u00b7 GGUF Q4_K_M", |
| "agree": 91.4, |
| "n": 500, |
| "size_gb": 1.27, |
| "size_note": "as reported on card", |
| "reference": "CUDA merged BF16", |
| "source": "https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-v2-GGUF/raw/main/README.md" |
| }, |
| { |
| "tier": "4-bit", |
| "model": "v1", |
| "label": "Jev-Style 2B v1 \u00b7 GGUF Q4_K_M", |
| "agree": 94.39999999999999, |
| "n": 500, |
| "size_gb": 1.3, |
| "size_note": "as reported on card", |
| "reference": "bf16", |
| "source": "https://huggingface.co/chaoliangUNSW/Jev-Style-Qwen3.5-2B-Decision-GGUF/raw/main/README.md" |
| } |
| ], |
| "deltas": { |
| "v2 Q4_K_M size / v3 Q4_K_M size": 2.4 |
| }, |
| "no_agreement_delta_claimed": "v3 parity rows are training-pool rows; v1/v2 used 500 held-out decisions, so no agreement gap is claimed", |
| "long_points": "v3 formats also 100% top-1 at 16,384 and 25,600 tokens (3 rows each), parity report long_points", |
| "not_plotted": "mlx-4bit (98.75%, 237/240) is not a published format and is omitted", |
| "footnote": "v3: top-1 agreement with the PyTorch FP32 reference on a 240-row parity fixture drawn from the training pool (22 categories, en+zh), plus 6 extra rows at about 16K and 25.6K tokens (6/6 agree). v3 sizes = exported weight files (GB = 10^9 bytes). 2B v1/v2: as reported on their public HF GGUF cards (500 held-out decisions each; vs bf16 for v1, vs CUDA merged BF16 for v2; card sizes). Different fixtures (training-pool rows for v3, held-out rows for v1/v2) and references: rows are not a paired comparison. x-axis starts at 88%.", |
| "protocol_label": "v3 G5 parity: 240-row mixed fixture (drawn from training-pool rows, 22 categories, en+zh) vs torch FP32; plus 16K and 25.6K long points (3 rows each). 2B numbers are from their cards on different fixtures (500 decisions vs bf16)" |
| } |
