SmolCUA — Computer-Use Aptitude

SmolCUA evaluates text-based computer-use reasoning: tool selection, next actions, strategy switching, instruction following, ambiguity, coding, structured output, coordinates, multi-constraint formatting, and state tracking. It is a synthetic text benchmark, not live UI automation or a visual computer-use environment.

400 questions per pass, three passes, 1,200 total. Each question contains four options and an explicit response contract. Models generate the answer; there is no forced-choice logit shortcut, output repair, or answer extraction.

Current edition

The default is compact v1.4, with 400 item positions across three passes, 40 items from each of ten categories per pass, for 1,200 total evaluations. It retains passes 1–3 of v1.3 exactly: questions, options, answer contracts, keys and weights are unchanged. Each pass keeps a 100-point category-weighted budget. The mean score and worst-pass reliability are computed over these three passes. Run python make_v14.py to reproduce the edition from bundled v1.3.

The 400 positions were selected in v1.3 using fixed SHA256 ranking, four positions per category/response-format stratum in pass 1. The same positions recur across passes. The Hugging Face default configuration and runner/scorer defaults now use v1.4. The earlier 2,000-instance, five-pass v1.3 remains available through compact_v1_3 and --run-subdir runs/v1_3_seed_20261023; the full 5,000-instance v1.2 remains available through corrected_v1_2 and --run-subdir runs/v1_2_seed_20261023. Use --run-dir for a local edition path.

Scores and worst-pass reliability from three-pass and five-pass editions are not interchangeable. Report the edition, question hash and scoring protocol.

v1.4 question SHA256: 74fb0a12a8dd75cbd8675d76e8bcf4a16e1b424668a195f4ff9a31da2becf065. The manifest records selected item IDs, source and scoring hashes, category budgets, and the selection procedure.

The source v1.2 inherits v1.1's 250 repairs and supplies missing original target coordinates in 32 additional scaled-display questions. Missing coordinates are reconstructed during dataset authoring from the keyed result divided by the stated scale. This creates a consistent task, not a claim to have recovered missing original generator inputs. Each edit is recorded in runs/v1_2_seed_20261023/corrections.json. Inference never reads answer keys to construct prompts or select candidates.

Scoring ignores surrounding whitespace. A response such as \n A \n earns the same credit as A. Whitespace inside payloads remains meaningful (for example, TSV separators or spaces inside literal strings). Explanations and code fences still violate contracts that prohibit them. An additional diagnostic counts correct answers after removing an outer code fence; it is not the primary score. The scoring protocol is recorded as v2-outer-whitespace-insensitive; scores must identify both dataset version/hash and scoring protocol.

Chat templates default to enable_thinking=False; prompts explicitly ask for candidate selection without real tool execution or extra explanation. Token-limit stops and format failures are reported. Use a fresh output directory when changing these settings. The bundled data is preferred over a download unless a remote revision is explicitly requested.

Run

Download this repository's toolkit files, then:

pip install -r requirements-transformers.txt
python run_smolcua.py --model YOUR_ORG/YOUR_MODEL --backend transformers \
  --dataset dotlabs/smolcua --prompt-mode auto --max-model-len 4096 \
  --output-dir results/my-model --score

plain is appropriate for base models and provides zero-shot prompts. Use --prompt-mode auto or chat for instruction-tuned models with chat templates. There are no few-shot exemplars in either mode. --dataset-revision COMMIT pins the dataset; --revision COMMIT pins model/tokenizer loading.

The repository is a public self-scoring edition, including files under private/ for evaluation. The directory name does not make them confidential. It is not a blind leaderboard. Disclose exposure to answer keys when reporting results; do not use test questions/keys for training.

Pretokenization and inference

• All selected prompts are compiled in CPU batches before Transformers model weights are loaded. The cache key includes actual prompt contents, dataset identity, tokenizer vocabulary/configuration, and prompt formatting.
• Cache hits reuse exact input IDs. No prompt truncation is allowed.
• Transformers uses length/budget buckets, greedy cached decoding, a default batch size of 32, and CUDA OOM backoff down to one. Supported models compute only the last-position logits during generation.
• Tokenization, model loading, prefill, decode tokens, completed questions, elapsed time, throughput, and ETA appear in flushed progress logs. Long calls get heartbeat updates; no fabricated completion percentage is shown.
• EOS counts as a generated token; padding after a completed sequence does not. Raw response text is not printed in progress logs.
• vLLM supports pretokenized inputs and its native generation progress. API mode uses concurrent requests and server-side tokenization.

# Compile prompts without allocating model weights:
python run_smolcua.py --model YOUR_MODEL --pretokenize-only --prompt-mode plain

# Larger batches; CUDA OOM backoff preserves completed answers:
python run_smolcua.py --model YOUR_MODEL --backend transformers \
  --batch-size 64 --progress-interval 5 --output-dir results/model --score

Performance depends on model, hardware, padding, and output lengths. Batched floating-point inference can differ from single-item inference near logit ties; record the batch size and precision. This release does not claim a measured GPU speedup. See RUNNING_AND_PUBLISHING.md for settings.

Scores and outputs

• Mean weighted score (0–100), with category-specific weights.
• Worst-pass exact-answer reliability: 100 * min(correct per pass) / 400.
• Per-pass, category, response-format, and consistency breakdowns.
• Raw answers, dataset hash/edition, configuration, token-cache statistics, actual token counts, generation batching statistics, and throughput.

Output files are predictions.jsonl, run_metadata.json, performance.json, and score.json. --score requires all 1,200 responses in the default edition (all instances of the selected edition); partial requests are rejected before model loading. The standalone scorer can assess missing answers as incorrect. Strict JSON contracts reject duplicate keys.

Weights are research-informed and have not been calibrated against a large representative model sample. v1.1 retains v1.0 weights despite clearer wording. Reference-answer and empty-answer tests validate scorer behavior; they are not model performance results.