Identify the tested artifact
A model family is a useful starting filter, but it is not a complete model identity. Preserve the developer, release, parameter variant and any fine-tune or other modification reported by the author. For a reproducible comparison, record the repository, exact file and revision when available. Hugging Face supports downloads pinned to a commit, so a repository name alone need not be the final level of detail.
On AMD-Pro-Bench, follow the model reference and the source report. A link to a base model is useful context, but it should not be treated as confirmation that those exact weights were tested. Leave unresolved artifact details explicit.
Keep the full quantization name
Quantization stores values at lower precision to reduce memory requirements, with an accuracy trade-off that depends on the method and task. Nominal bit depth does not identify the full scheme. GGUF supports distinct formats with different block layouts and metadata overhead; several formats can share a similar bit depth without being interchangeable.
Use the exact reported label, rather than merging everything into 3-bit or 4-bit. Record weight precision separately from KV-cache precision. Treat memory use, speed and evaluation quality as separate outcomes: a smaller representation does not by itself prove faster inference on a particular AMD card and backend.
Hold the evaluation conditions steady
To investigate a quantization trade-off, first hold the underlying checkpoint constant. Then match the benchmark version, test set, harness and scoring definition. Check prompt formatting, few-shot examples, generation settings and sample limits as well. The LM Evaluation Harness exposes these as distinct options; sharing a benchmark family name does not establish that the runs used identical conditions.
Use the GPU and model filters to find candidate pairs, then verify them in their original reports. If two rows use different fine-tunes or test subsets, describe them as different configurations. Do not attribute their score gap solely to quantization. When no comparable quality evaluation is reported, the database can document the measured speed while leaving the quality trade-off unanswered.