---
license: apache-2.0
tags:
- benchmark
- quantization
- gptq
- awq
---

# QuantBench leaderboard data

Raw benchmark data behind the [QuantBench leaderboard](https://quantbench.pages.dev):
calibration-quality GPTQ/AWQ quantization results across model sizes, calibration
corpora, and GPU tiers. **331 rows** (239 ok / 92 failed — failed
runs are published too; a documented failure is a finding, not noise).

## Models

- `Qwen/Qwen2.5-1.5B-Instruct` (1.5B)
- `HuggingFaceTB/SmolLM2-1.7B-Instruct` (1.7B)
- `deepgrove/Bonsai` (0.5B)
- `Qwen/Qwen2.5-3B-Instruct` (3B) — licence pending, rows only, no weights here
- `HuggingFaceTB/SmolLM3-3B` (3B)

## Files

- `rows.csv` — the full table (see the leaderboard site for column definitions
  and the methodology page for the eval protocol).
- `<row_id>.json` — one file per row, same fields as its `rows.csv` line.
- `niche_rows.csv` — task-level niche columns over the same artifacts
  (HumanEval[0:50] pass@1, FiQA-2018 sentiment accuracy [flare-finqa is gated
  for this token — instrument recorded per row], SEC-filing context drift,
  10-prompt ROUGE-L consistency), each vs a same-stack fp16 baseline;
  failures/skips recorded as rows.
- `logs/` — the driver log, last heartbeat, and spend ledger from the run that
  produced this data (transparency, not polish — these are raw operational logs).
- `pools/BUILD.md` — exact recipe to reproduce the calibration/eval text pools.
  **No third-party corpus text is redistributed here.**

## Honesty notes carried into every row

- GPTQ rows in this sweep ran gptqmodel's torch-fallback kernel path
  (`load_backend` column), not the optimized kernel — perplexity is
  unaffected, throughput numbers understate what a working kernel would show.
- fp16 baselines are per-stack (autoawq vs gptqmodel use different
  torch/transformers pins), so `ppl_delta_vs_fp16` is only comparable within
  one stack and one GPU tier, never across them.
- `webgpu_runnable` is `[UNVERIFIED]` on every row — no browser measurement
  was run.
- Machine-generated benchmark output. Independent verification welcome;
  treat any single row as a data point, not a certified result.