--- pretty_name: CogniRank reproducible evaluator language: - en tags: - evaluation - reasoning - small-language-models --- # CogniRank: independent reproduction This public dataset repository distributes **evaluation software**, including a single self-contained `cognirank.py`. No Modal account or original project checkout is needed. The November update adds SmolScript at 20%, using **350 fixed examples by default across nine programming languages**. The script embeds the five benchmark implementations, SmolCUA questions and answer keys, the fixed SmolScript subset and execution sandbox, source checksums, and selected Procedural Pile row indices. Internet access is needed for Python packages, model weights, and the other pinned datasets. The evaluator, source manifest, and published measurements for this update are frozen at [dataset commit `5f5969a2bae453e45f2a51c746908e16b21cc2ab`](https://huggingface.co/datasets/dotlabs/cognirank/tree/5f5969a2bae453e45f2a51c746908e16b21cc2ab). The command below downloads the current evaluator, including compatibility fixes; use that frozen commit to reproduce the original software exactly. ## Run Use Linux, Python 3.11, git, a C/C++ build toolchain, Docker Engine, and an NVIDIA GPU supporting CUDA and bfloat16, with enough GPU memory for your model and the fixed batch sizes. Allow up to 16 GB of host RAM for eight concurrent execution containers. Download `cognirank.py` from this public repository. Authenticate with Hugging Face only if a selected model requires access: ```sh curl -L https://huggingface.co/datasets/dotlabs/cognirank/resolve/main/cognirank.py -o cognirank.py python3.11 -m venv .venv . .venv/bin/activate python -m pip install --upgrade pip python cognirank.py --verify python cognirank.py --install-deps python cognirank.py --build-sandbox python cognirank.py --model dotlabs/void.1 --output-dir void1-results ``` Multiple models run sequentially: ```sh python cognirank.py --models dotlabs/void.1 MinimaLabs/KLM-1.5 ibm-granite/granite-4.0-h-350m --output-dir comparison python cognirank.py --models dotlabs/void.1 MinimaLabs/KLM-1.5 ibm-granite/granite-4.0-h-350m --output-dir comparison --resume ``` `--dry-run` prints the protocol and model revisions without installing dependencies or running inference. `--extract-only ./cognirank-source` extracts the complete bundled source for inspection. `--print-requirements` prints dependencies. Installation explicitly executes pip in the current environment; inference permits Hugging Face custom model code, as the original evaluator does. Known model revisions are pinned in `source_manifest.json`. Other model IDs resolve `main` once and record the immutable revision. To choose a different revision for one model, use `--model-revision COMMIT_SHA`. Use a new output directory for a different protocol, revision, or context. Resume preserves the resolved revision and verifies completed component checksums; SmolCUA resumes individual predictions, and Procedural Pile can re-score saved generations. Interrupted MARC/TTM components restart. SmolScript reuses a fully saved generation checkpoint and repeats unfinished component scoring. ## Exact protocol | Component | Examples | Weight | Metric | |---|---:|---:|---| | SmolCUA v1.4 | 1,200: three fixed passes of 400 | 24/95 | Mean weighted pass score, 0–100 | | MARC | 320: 4 families × 4 levels × 20 | 20/95 | Mean exact-match accuracy of MARC-0…3 | | Procedural Pile | Fixed 350 across all 55 task families | 20/95 | Mean official verifier reward × 100 | | Tiny Theory of Mind | All 2,000 | 12/95 | Token-normalized likelihood multiple-choice accuracy | | SmolScript | Fixed 350 across nine languages | 19/95 = 20% | Macro-average of nine language Pass@1 accuracies | `CogniScore = (24*SmolCUA + 20*MARC + 20*ProceduralPile + 12*TTM + 19*SmolScript)/95`. All components are required. Unsupported context is reported per benchmark; completed component scores are retained, the overall index is omitted, and the model is excluded from the ranking. Other failures still stop the run. The rounded display weights are 25.26%, 21.05%, 21.05%, 12.63%, 20.00%; computation uses the exact fractions above. Generation is greedy, bfloat16, CUDA device 0. SmolCUA uses automatic chat/plain formatting, batch 32, a 512-token response cap (48 for short formats). MARC uses the upstream plain demonstration prompts, deterministic seed string `0`, batch 32, and 200 generated tokens. Procedural Pile uses automatic chat formatting with thinking disabled, batch 8, and 512 generated tokens. TTM uses plain continuations and `lm-eval==0.4.12`, batch 8. SmolScript uses plain code-completion prompts for base and instruction models alike (Python prefix `def solve(n):\n`, without trailing indentation), batch 64, one greedy continuation, and language/difficulty budgets of 96–256 tokens. Hidden tests determine correctness by execution or DOM/computed-style assertions. Equivalent behavior passes regardless of code text; whitespace is preserved where it changes program or string semantics. Partial hidden-test pass rate is reported separately and does not replace task Pass@1. Full implementations are embedded; prompt construction, whitespace-aware scoring, and selection indices can be inspected with `--extract-only`. Default requested evaluation context is **4,096**. Extending a model's declared context is recorded as a context-extension experiment. Known fixed learned-position architectures are capped at their hard positional limit. Response budgets shrink to remaining space; input prompts are never truncated. A prompt that cannot fit produces an explicit unsupported-context report. `--eval-context 0` uses the native context, capped at 4,096 by this runner; changing context changes the experiment. Effective limits and shortened response budgets are recorded in component outputs. Procedural Pile uses the pinned official verifier with `inverse-math-normalize-progress-v1`: it rejects a non-progressing malformed LaTeX fraction instead of hanging indefinitely. Other verifier timeouts fail the evaluation rather than silently assigning zero. The patch and upstream license are embedded in `inverse_math_normalizer.py`. ## Immutable sources - **SmolCUA:** [`dotlabs/smolcua`](https://huggingface.co/datasets/dotlabs/smolcua/tree/faf2b33f893573e23666a4e59c99a09d0b73ba51), revision `faf2b33f893573e23666a4e59c99a09d0b73ba51`, `runs/v1_4_seed_20261023/`. Bundled questions SHA256: `74fb0a12a8dd75cbd8675d76e8bcf4a16e1b424668a195f4ff9a31da2becf065`. - **MARC generator:** [`FlameF0X/Micro-ARC/marc.py`](https://huggingface.co/spaces/FlameF0X/Micro-ARC/resolve/4ed8c708aa3e9665a9d9c81f2e3bc586cf60e14e/marc.py), revision `4ed8c708aa3e9665a9d9c81f2e3bc586cf60e14e`. - **Procedural Pile:** [`reasoning-core/procedural-pile`](https://huggingface.co/datasets/reasoning-core/procedural-pile/resolve/0e3216537b0cb7df8709c6cae34bfceae6450068/data/test-00000.parquet), revision `0e3216537b0cb7df8709c6cae34bfceae6450068`. Full test shard SHA256: `4ed623b9ad54a2f08483c8281825000e2f46c1e8ccfdeddd1db1cb1ca5c6cac8`. The 350 selected IDs and family counts are embedded in `source_manifest.json`; selection version `procedural-pile-350-v1`, subset SHA256 `67647511ff184a44c9f1cb7abaed66e0cd113f2deecafba95250b9a7338ddf6f`. Twenty families contribute seven rows and 35 contribute six; exact hash-based selection is in the bundled parser. - **Official verifier:** [`sileod/reasoning_core`](https://github.com/sileod/reasoning_core/tree/68b0b3a0f29128d2c56b3eafcdaa34777b9e1279), revision `68b0b3a0f29128d2c56b3eafcdaa34777b9e1279`. - **TTM:** [`AxiomicLabs/Tiny_Theory_of_Mind`](https://huggingface.co/datasets/AxiomicLabs/Tiny_Theory_of_Mind/resolve/ba4c644ea7ace67e096305fe2c9f87c4cf4879ad/tiny_theory_of_mind_2000.jsonl), revision `ba4c644ea7ace67e096305fe2c9f87c4cf4879ad`, SHA256 `58827a0361597ae4d5ffa79ddbab0936d1509cafcfc02cbb731e4d7a7f3584b8`. - **SmolScript:** [`dotlabs/smolscript`](https://huggingface.co/datasets/dotlabs/smolscript/tree/fa26615d2eb867def83882444324babb6eb2f4e6), revision `fa26615d2eb867def83882444324babb6eb2f4e6`; fixed dataset selection `smolscript-0.3.0`, evaluation protocol `smolscript-0.5.0`, prompt version `smolscript-base-completion-v3`. The bundled manifest lists all 350 IDs, language/difficulty allocations, and data/test SHA256 checksums. - **KronLSTM dependency:** [`NX-AI/KronLSTM`](https://github.com/NX-AI/KronLSTM/tree/3d800c9a153a88ee9624fe6acf7f39a58017a17e), installed without dependency replacement after the core requirements. The source manifest includes all 15 published model repository names and exact revisions, every bundled source checksum, and source download URLs. Save the Hugging Face commit of this release and `SHA256SUMS` alongside your results. ## Execution sandbox `--build-sandbox` builds the embedded Dockerfile once with internet access. It uses Playwright Python `v1.56.0-noble`, Chromium, GCC/G++, Node, Lua 5.4, Mono C#, Rust, html5lib 1.1, and tinycss2 1.4.0. Distribution runtime packages resolve at image build time; keep the recorded image ID or archive the image for byte-identical reruns. Browser/runtime preflight runs before SmolScript generation. Each candidate executes in a fresh credential-free container with network disabled, a read-only root filesystem, an unprivileged UID, dropped capabilities, 2 GB memory, one CPU, a 512-process limit, and phase-specific timeouts. No candidate code executes on the inference host. Infrastructure failures stop evaluation without assigning a zero. References and hidden expectations stay outside candidate containers; only fixtures and input values cross the boundary. Evaluation does not need internet inside the sandbox. The single-file evaluator needs no Modal account. Its local Docker runner uses the same prompts, generation implementation, hidden tests, source builders, browser assertions, and scoring as the hosted run. Container scheduling and image builds differ; runtime identities and preflight output are recorded. ## Outputs and reproducibility limits Each model directory contains `run.json`, `results.json`, per-component logs and predictions, `environment.json`, and `pip-freeze.txt`. `comparison.json` contains the completed model scores. `published_results.json` records the 15 published checkpoint revisions, component scores, updated indices, and original result checksums. SmolScript also records per-example outcomes, per-language accuracy, syntax success, test pass rate, and its sandbox image identity. The release identifier hashes bundled source files; resume rejects a different release. The single-file archive is checksum-verified before extraction. This release implements the **November CogniRank update**. Previous four-component aggregate scores are superseded; unchanged component measurements are reused only at the same model revision and evaluation context. The published snapshot combines those measurements with completed SmolScript 0.5.0 runs, and records the source result checksums. Older SmolScript protocols are excluded. Core inference versions are pinned; some secondary dependency ranges remain flexible and the actual environment is recorded. GPU/kernel differences can affect numerical results. The portable packaging and score recomputation were checked locally. A complete GPU benchmark has not been rerun with this distribution, and the full SmolScript reference suite remains execution-unverified; see its bundled validation report. Published model runs are completed model evaluations, not a claim that every reference solution has been validated. Source datasets retain their upstream licenses and attribution. SmolCUA's source README and the inverse-math patch's MIT notice are included in the archive. ## Model compatibility update (November 10) The shared generation and likelihood loaders adapt legacy Transformers tied-weight lists using verified parameter aliases. Inferno and Min-Spark native context limits are enforced before generation, preventing upstream silent sliding-window truncation. Min-Spark models declare a 512-token context; benchmarks that cannot fit report unsupported context and omit the overall index. This does not extend their trained context. `compatibility_targets.json` records the top ten models for each Open SLM Leaderboard size filter (3M, 10M, 50M, 100M, 150M, 250M, and all sub-400M), plus explicitly requested variants. It records configuration inspections and pinned model revisions. These are compatibility targets, **not a claim of successful execution**: full checkpoint/GPU smoke tests remain outstanding. The alias adapter has local regression coverage; the Inferno checkpoint has not been rerun here.