| { |
| "tag": "nvfp4", |
| "meta": { |
| "harness": "https://git.srk.rest/shadowrock/uocr-quant @ 1f584d7", |
| "date": "2026-08-17", |
| "base_model": "baidu/Unlimited-OCR", |
| "base_model_revision": "07dea832e22aefee32ad281d4b80551282e1c168", |
| "engine": "vLLM 0.26.0 (enforce_eager, max_model_len 8192)", |
| "torch": "2.11.0+cu130", |
| "gpu": "NVIDIA GeForce RTX 5070 Ti (SM120), driver 610.43.03", |
| "decoding": "greedy (temperature 0.0), max_tokens 1024", |
| "cer_definition": "Levenshtein/len(ref) on whitespace-normalized text after stripping grounding/box tags, vs BF16 reference transcript" |
| }, |
| "model_dir": "exports/Unlimited-OCR-NVFP4", |
| "prompt": "<image>\n<|grounding|>OCR this image.", |
| "decode_tok_per_s": 40.84, |
| "peak_vram_gib_client": null, |
| "fixtures": { |
| "invoice.png": { |
| "raw": "title [54, 79, 402, 127]INVOICE #2026-0817\ntext [54, 258, 456, 308]Bill To: ShadowRock LLC\ntext [54, 354, 577, 410]Item: GPU Quantization Services\ntext [54, 443, 402, 498]Qty: 2 Unit: $150.00\ntext [54, 625, 356, 672]Total Due: $300.00\ntext [54, 720, 527, 768]Due Date: September 1, 2026", |
| "ntok": 145, |
| "seconds": 6.129 |
| }, |
| "memo.png": { |
| "raw": "title [37, 47, 304, 87]MEMORANDUM\ntext [37, 144, 305, 180]To: Engineering Team\ntext [37, 211, 253, 247]From: Operations\ntext [37, 280, 310, 318]Date: August 17, 2026\ntext [37, 351, 501, 387]Re: GPU cluster maintenance window\ntext [34, 488, 545, 531]The primary GPU cluster will be offline for\ntext [34, 562, 639, 601]scheduled maintenance from 02:00 to 06:00 UTC\ntext [34, 628, 567, 670]on August 24. Batch jobs should be drained\ntext [34, 699, 609, 737]before 01:30 UTC. Contact ops@example.com\ntext [34, 768, 270, 807]with any questions.", |
| "ntok": 234, |
| "seconds": 3.89 |
| }, |
| "table_report.png": { |
| "raw": "txt\ntitle [58, 52, 521, 103]Quarterly Revenue Report\ntable [58, 158, 942, 454]<table>QuarterRevenueGrowthQ1 2026$4.2M12%Q2 2026$4.9M16%Q3 2026$5.1M4%</table>\ntext [53, 505, 681, 602]Operating expenses decreased 3% QoQ, driven by reduced cloud infrastructure spending.", |
| "ntok": 146, |
| "seconds": 2.836 |
| } |
| }, |
| "cer_vs_bf16": { |
| "invoice.png": 0.0, |
| "memo.png": 0.0, |
| "table_report.png": 0.0204 |
| }, |
| "mean_cer_vs_bf16": 0.0068 |
| } |
