| --- |
| license: mit |
| language: |
| - en |
| tags: |
| - tool-use |
| - function-calling |
| - mcp |
| - model-context-protocol |
| - benchmark |
| - quantization |
| - leaderboard |
| pretty_name: QuantMCP Results |
| --- |
| |
| # QuantMCP Results |
|
|
| Real benchmark results for the |
| [QuantMCP benchmark](https://github.com/Happynood/quant-mcp-bench), measuring |
| how quantization degrades LLM function-calling reliability against real, |
| unmodified [Model Context Protocol](https://modelcontextprotocol.io) server |
| tool schemas — and whether that degradation matches what |
| [QuantCall](https://huggingface.co/datasets/happynood/quantcall-results) |
| already measured on curated BFCL/ToolACE schemas. Every row comes from an |
| actual `quantmcp run` execution against a real, sandboxed MCP server — no |
| fabricated or hand-edited numbers. |
|
|
| ## Headline finding: Cross-Benchmark Consistency (CBC) |
|
|
| **CBC (Spearman rho) = -0.755 (n=8 model×quant pairs, 3 model families)** — |
| see `data/cbc.json`. |
|
|
| QuantCall's BFCL-measured quantization degradation does **not** reliably |
| carry over to real MCP schemas: the correlation is negative (degradation |
| directions often flip), and it got *more* negative, not less, as more data |
| was added. With 2 model families (Qwen3-0.6B, Llama-3.2-1B) the estimate |
| was -0.551 (n=6) — itself the product of three increasingly-averaged |
| computations that ranged -0.824 to -0.265 before stabilizing. Adding |
| Qwen3-1.7B as a 3rd family (a real within-family size contrast against |
| Qwen3-0.6B) moved it to -0.755 (n=8): the sign held, the magnitude |
| strengthened. See |
| [`docs/RUN_REAL.md`](https://github.com/Happynood/quant-mcp-bench/blob/main/docs/RUN_REAL.md) |
| in the GitHub repo for the full convergence table and honesty caveats — |
| n=8 is still far too few for a rigorous p-value on a Spearman correlation. |
|
|
| ## Files |
|
|
| | File | Grain | Description | |
| |------|-------|--------------| |
| | `data/raw_results/**/*.result.json` + `*.manifest.json` | one pair per real run | Every real run this project has produced (96 result files across 3 model families × 4 server tiers, with varying quant/repeat coverage per family — Qwen3-0.6B and Llama-3.2-1B at 4 quant levels with 3 independent repeats on tiers U1-U3, Qwen3-1.7B at 3 quant levels, single run, no fp16 (its bf16 weights don't fit a 4GB card at a usable context length — see `docs/RUN_REAL.md`)), each with a full manifest (git SHA, config/fixture hashes, hardware fingerprint). The `model` field is a portable `~/models/...` path, not a specific machine's absolute path. Since Phase 7, each file also carries a `instances` array (one entry per task instance, tagged with the tool it targeted) enabling the per-tool SCI regression below. | |
| | `data/mcp_runs.csv` | one row per real run | Flattened, path-sanitized view of every `raw_results` file | |
| | `data/mcp_tier_breakdown.csv` | one row per server tier | Mean SVR-MCP/TSR/η per tier, annotated with that tier's real Schema Complexity Index (SCI) | |
| | `data/cbc.json` | one row per (model, quant) pair | Cross-Benchmark Consistency deltas against QuantCall's published BFCL numbers, plus the Spearman rho itself | |
| | `data/sci_regression.json` | one row per live tool | Per-tool Schema Complexity Index (SCI) paired with its own Δ SVR-MCP (fp16 vs. Q4_K_M), plus an OLS slope and bootstrap 95% CI across all 38 covered tools — the statistically-powered version of the SCI-vs-degradation question that `mcp_tier_breakdown.csv`'s 4-tier view alone can't answer | |
|
|
| ## Schema: `data/mcp_runs.csv` |
| |
| | Column | Type | Description | |
| |--------|------|--------------| |
| | `model` | string | Sanitized model name (local GGUF paths stripped to a canonical name — see `report/published.py::sanitize_model_name` in the GitHub repo) | |
| | `quant` | string | Quantization level: fp16, Q8_0, Q5_K_M, Q4_K_M (Qwen3-1.7B: Q8_0/Q5_K_M/Q4_K_M only) | |
| | `tier` | string | MCP server tier: filesystem, git, sqlite, or memory | |
| | `n` | int | Number of task instances evaluated | |
| | `svr_mcp` | float | Schema-Validity Rate against the real, live tool schema (SVR-MCP, spec §4.1) | |
| | `tsr` | float | Task Success Rate — the call was schema-valid *and* produced the correct outcome (spec §4.2) | |
| | `vram_gb` | float | Peak VRAM usage in GB for this run | |
| | `eta` | float | Reliability-per-VRAM: `(0.5*svr_mcp + 0.5*tsr) / vram_gb` | |
| | `pareto_optimal` | bool | Whether this (model, quant, tier) config sits on the reliability-vs-VRAM Pareto frontier | |
|
|
| ## Schema: `data/cbc.json` |
|
|
| ```json |
| { |
| "rho": -0.755, |
| "n_pairs": 8, |
| "table": [ |
| {"model": "...", "quant": "...", "baseline_quant": "...", "delta_svr_bfcl": ..., "delta_svr_mcp": ...} |
| ] |
| } |
| ``` |
|
|
| `delta_svr_bfcl` is QuantCall's published BFCL SVR delta vs. that model's |
| own baseline quant (fp16 for every family except Qwen3-1.7B, which uses |
| Q8_0 — see `baseline_quant` per row); `delta_svr_mcp` is this project's |
| equivalent delta on real MCP schemas (pooled across all four server tiers, |
| weighted by task count). |
|
|
| ## Schema: `data/sci_regression.json` |
| |
| ```json |
| { |
| "n": 38, |
| "slope": 0.045, |
| "intercept": ..., |
| "slope_ci": [-0.064, 0.170], |
| "points": [ |
| {"tool": "...", "tier": "...", "sci": ..., "delta_svr": ..., "n_baseline": ..., "n_quant": ...} |
| ] |
| } |
| ``` |
| |
| `slope`/`slope_ci` describe the OLS fit of Δ SVR-MCP (fp16 minus Q4_K_M |
| pass rate, pooled across model families weighted by n) against each live |
| tool's own SCI. The 95% CI is a percentile bootstrap over the (SCI, Δ) |
| pairs, not a parametric estimate. |
|
|
| ## How to Submit |
|
|
| 1. Run the benchmark on your hardware following |
| [docs/RUN_REAL.md](https://github.com/Happynood/quant-mcp-bench/blob/main/docs/RUN_REAL.md). |
| 2. Verify your `result.json` contains a `manifest` block with a git SHA and |
| fixture hash. |
| 3. Open a PR on [GitHub](https://github.com/Happynood/quant-mcp-bench) adding |
| your result file under `results/`. |
| 4. Run `quantmcp leaderboard results/ --output-dir leaderboard/` and include |
| the regenerated CSVs in your PR. |
|
|
| ## Links |
|
|
| - GitHub: https://github.com/Happynood/quant-mcp-bench |
| - Eval suite: https://huggingface.co/datasets/happynood/quantmcp-suite |
| - Leaderboard (Space): https://huggingface.co/spaces/happynood/quantmcp-leaderboard |
| - Sibling project (curated-schema quantization benchmark): https://github.com/Happynood/quant-toolcall-bench |
|
|
