README.md

metadata

license: apache-2.0
language:
  - ar
  - en
task_categories:
  - text-generation
tags:
  - code
  - arabic
  - python
  - verifiable-rewards
  - grpo
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train.jsonl
      - split: test
        path: data/test.jsonl

Arabic Code Packet A11 (verifiable)

Arabic-instruction Python tasks where every answer is machine-checkable. Built by ABUZ8 LLC for the SUPERBRAIN-1 Arabic brain (https://huggingface.co/abuz8studios/SUPERBRAIN-1).

• 6,021 train / 1,026 test, 27 deterministic template generators, seed 11.
• No LLM-generated content: every prompt, reference solution and test comes from code in tools/build_packet_a11.py.

Kind	Train	Test	How it is graded
predict-the-output	2,230	380	exact match with the snippet's real stdout (executed in a subprocess at build time)
write-a-function	2,453	418	5-8 hidden unit tests
fix-the-bug	669	114	hidden tests; buggy version verified to fail
complete-the-function	669	114	hidden tests

Verification (tools/check_a11.py, exit 0)

[counts] train=6021 exam=1026 generators train=27 exam=27 missing_fields=0 dup_prompts=0 -> OK
[reference] sample=400 pass=400/400 (100.00%) | fix_bug buggy versions passing=0/67 -> OK
[predict] sample=200 re-executed_match=200/200 -> OK
[negative] other-item completions scored correct: 0/100 -> OK
[leakage] train_vs_exam exact 0, norm 0, qnorm 0, qhash 0, code 0 -> LEAKAGE=0

Test items use a phrasing template never seen in train and parameter combinations disjoint from train.

Fields

task, src, kind, topic, template, prompt, completion, answer, qhash, lang; predict items add code; code items add entry, reference, tests, test_code (fix-the-bug also buggy).

Scoring

python tools/a11_score.py ITEMS.jsonl PREDS.jsonl - runs model code against hidden tests in a subprocess with a 5 s timeout (process isolation, not a security sandbox).

Model scores

None yet. Baselines for SUPERBRAIN-1 will be added here once measured.