| --- |
| license: mit |
| language: |
| - en |
| tags: |
| - tool-use |
| - function-calling |
| - benchmark |
| - quantization |
| - leaderboard |
| pretty_name: QuantCall Results |
| --- |
| |
| # QuantCall Results |
|
|
| Real benchmark results for the |
| [QuantCall benchmark](https://github.com/Happynood/quant-toolcall-bench), |
| measuring how quantization degrades LLM function-calling reliability. |
| Every row comes from an actual `quantcall run` execution — no fabricated or |
| hand-edited numbers. |
|
|
| ## Files |
|
|
| | File | Grain | Description | |
| |------|-------|--------------| |
| | `data/runs.csv` | one row per real run (per seed) | Raw per-seed data with full manifest (git SHA, config/dataset hashes) | |
| | `data/leaderboard.csv` | one row per (model, quant, backend, decoding, tier) | Aggregated over seeds, with bootstrap 95% CIs and deltas vs an explicit baseline quant | |
|
|
| ## Schema: `data/runs.csv` |
|
|
| | Column | Type | Description | |
| |--------|------|-------------| |
| | `model` | string | Model identifier (HF repo ID or local path) | |
| | `quant` | string | Quantization level: fp16, Q8_0, Q5_K_M, Q4_K_M, AWQ, GPTQ | |
| | `backend` | string | Inference backend: llama-cpp, transformers, vllm, openai | |
| | `decoding` | string | Decoding mode: free or constrained | |
| | `tier` | string | Dataset tier(s) evaluated, `+`-joined (e.g. `T1+T6`) | |
| | `seed` | int | Random seed for this run | |
| | `sample_size` | int | Number of instances evaluated per tier | |
| | `svr` | float | Schema-Validity Rate [0, 1] | |
| | `tsa` | float | Tool-Selection Accuracy [0, 1] | |
| | `ac` | float | Argument Correctness [0, 1] | |
| | `abstention` | float | Abstention Accuracy [0, 1] | |
| | `fcr` | float | Function-Calling Reliability — 0.25 × (SVR + TSA + AC + Abst) | |
| | `vram_gb` | float | Peak VRAM usage in GB for this run (empty if not measured) | |
| | `git_commit` | string | QuantCall repo commit SHA used for this run | |
| | `config_sha256` | string | SHA-256 of the run config | |
| | `dataset_sha256` | string | SHA-256 of the evaluation sample | |
| | `timestamp` | string | ISO-8601 UTC timestamp of the run | |
|
|
| ## Schema: `data/leaderboard.csv` |
|
|
| | Column | Type | Description | |
| |--------|------|-------------| |
| | `model` | string | Model identifier | |
| | `quant` | string | Quantization level | |
| | `backend` | string | Inference backend | |
| | `decoding` | string | Decoding mode | |
| | `tier` | string | Dataset tier(s), `+`-joined | |
| | `n_seeds` | int | Number of seeds aggregated into this row | |
| | `fcr_mean` | float | Mean FCR across seeds | |
| | `fcr_ci_low` | float | Bootstrap 95% CI lower bound for FCR | |
| | `fcr_ci_high` | float | Bootstrap 95% CI upper bound for FCR | |
| | `svr_mean` | float | Mean SVR across seeds | |
| | `tsa_mean` | float | Mean TSA across seeds | |
| | `ac_mean` | float | Mean AC across seeds | |
| | `abstention_mean` | float | Mean Abstention across seeds | |
| | `vram_gb` | float | Mean peak VRAM in GB (empty if not measured) | |
| | `eta` | float | Efficiency: fcr_mean / vram_gb (empty if vram_gb is empty) | |
| | `delta_fcr_rel` | float | Relative FCR delta vs `baseline_quant` in the same scope; empty for the baseline row itself | |
| | `delta_ac_rel` | float | Relative AC delta vs `baseline_quant` | |
| | `baseline_quant` | string | The Δ reference quant for this scope — fp16 if it fits and was run, otherwise the best-available quant, always labeled explicitly here | |
|
|
| These two schemas are generated by `quantcall leaderboard <results_dir>` |
| (source of truth: `src/quantcall/report/published.py`, |
| `docs/RESULTS_SCHEMA.md` in the repo) — this card is kept in sync with that |
| code by a repo test (`test_no_schema_drift`). |
|
|
| ## How to Submit |
|
|
| 1. Run the benchmark on your hardware following [docs/RUN_REAL.md](https://github.com/Happynood/quant-toolcall-bench/blob/main/docs/RUN_REAL.md). |
| 2. Verify your `result.json` contains a `manifest` block with git SHA and hashes. |
| 3. Open a PR on [GitHub](https://github.com/Happynood/quant-toolcall-bench) adding your result file under `results/`. |
| 4. Run `quantcall leaderboard results/ --output-dir leaderboard/` and include the |
| regenerated `runs.csv` / `leaderboard.csv` in your PR. |
|
|
| ## Links |
|
|
| - GitHub: https://github.com/Happynood/quant-toolcall-bench |
| - This dataset: https://huggingface.co/datasets/happynood/quantcall-results |
| - Eval suite: https://huggingface.co/datasets/happynood/quantcall-suite |
| - Leaderboard (Space): https://huggingface.co/spaces/happynood/quantcall-leaderboard |
|
|
