task_id stringclasses 5
values | family stringclasses 3
values | split stringclasses 1
value | description stringclasses 5
values | reward stringclasses 3
values | sample_instances listlengths 2 2 |
|---|---|---|---|---|---|
numeric-math | numeric | train | Exact single-number math questions generated with DeepMind mathematics_dataset (linear equations, 2-variable systems, remainders, GCD, probability). | binary: 1 only for exactly {"value": <number>} equal to the answer at the requested precision | [
{
"seed": 1,
"files": {
"brief.md": "# Question\n\nLet b = 449407 + -449387. Let y = 3 + 6. Let w = -5 + y. Solve 2*s - 5*d = -3*s + b, -w*s + d = -4 for s.\n\n# What to submit\n\nSubmit exactly {\"value\": <number>}: the answer as a pure number, as an exact whole number.\nThe value must be a plain JS... |
media-m1 | media | train | Repair a video delivery timeline from metadata: reject clips for every applicable reason, exact frame arithmetic, gaps and overlaps. | binary: 1 only if the submitted object has exactly the requested keys and every value matches | [
{
"seed": 1,
"files": {
"brief.md": "# Media: repair a delivery timeline\n\nThese are synthetic video metadata, not video files. Read assets.json,\nedits.json and delivery.json. Process edits in their FILE ORDER, not ID order.\nAll frame numbers are integers; source ranges [in,out) exclude out.\nSourc... |
media-m2 | media | train | As media-m1 with differing frame rates and speeds, rational durations, gaps and overlaps. | binary: 1 only if the submitted object has exactly the requested keys and every value matches | [
{
"seed": 1,
"files": {
"brief.md": "# Media: repair a delivery timeline\n\nThese are synthetic video metadata, not video files. Read assets.json,\nedits.json and delivery.json. Process edits in their FILE ORDER, not ID order.\nAll frame numbers are integers; source ranges [in,out) exclude out.\nSourc... |
sec-path-v2 | security | train | Fix a path-traversal bug in a Flask download route while preserving legitimate behaviour (nested paths, Unicode, in-root symlinks). | binary: 1 only when every security and legitimate-use test passes on the submitted app.py, run in a restricted process | [
{
"seed": 1,
"files": {
"brief.md": "# Security repair\n\nRepair app.py while preserving the legitimate behavior below.\n\nRepair GET /download?file=<relative path>. The trusted root is app.config['FILE_ROOT']. Serve exact bytes with status 200 for regular files whose resolved path is inside that root... |
sec-idor-membership | security | train | Fix an insecure direct object reference in a Flask route: enforce authentication, tenant, archive and membership-based authorization. | binary: 1 only when every security and legitimate-use test passes on the submitted app.py, run in a restricted process | [
{
"seed": 1,
"files": {
"brief.md": "# Security repair\n\nRepair app.py while preserving the legitimate behavior below.\n\nRepair GET /boards?item=<record ID>. Trusted authentication is session['user_id']. It must be an exact Python int (not bool or a numeric string) identifying an enabled USERS entry... |
Evidence Desk
OpenEnv environment for the OpenEnv Arena. It covers three broad skills: exact multi-step math, reading structured files and applying every stated rule, and repairing real classes of web-security bugs without breaking legitimate behaviour. Every task uses the same tool interface and a strict, verifiable reward.
Tasks
| Task | Skill | What it tests |
|---|---|---|
numeric-math |
Math | Exact single-number questions: composed linear equations, 2-variable systems, remainders, GCD, probability |
media-m1 |
Rules over files | Repair a video delivery timeline from metadata: reject clips for every applicable reason, exact frame arithmetic |
media-m2 |
Rules over files | As m1 with differing frame rates and speeds, rational durations, gaps and overlaps |
sec-path-v2 |
Security | Fix path traversal in a Flask download route; keep nested paths, Unicode and in-root symlinks working |
sec-idor-membership |
Security | Fix an insecure direct object reference: authentication, tenant, archive and membership-based authorization |
Interaction
One JSON action per turn: open a file, calc (sum a list of numbers, or evaluate an
expression with + - * / **, common functions and pi/e), submit the answer object,
or finish. An optional work field holds scratch reasoning and is ignored by the
environment. Episodes end on submit, finish or after 14 calls.
Reward
Binary, with no partial credit.
- Math: 1 only for exactly
{"value": <number>}equal to the answer at the requested precision (exact answers; rounding ties to even). - Media: 1 only if the answer has exactly the requested keys and every value matches.
- Security: the model submits the complete repaired
app.py. It is imported in a restricted process (Landlock filesystem rules, a seccomp syscall allowlist, resource limits; no network, subprocesses or threads) and must pass every exploit and legitimate-use test. The tests run outside the candidate process.
Instances are generated fresh on each reset. tasks.jsonl lists each task with two
sample instances (visible files only); answers and tests are not published.
Data and licences
- Math questions (
data/deepmind_math_train.jsonl, 2,050 rows) were generated locally with DeepMind mathematics_dataset 1.0.1 (Apache-2.0, Copyright 2019 DeepMind Technologies Limited). Seedata/deepmind_math_LICENSE.txtanddata/deepmind_math_manifest.json. Held-out question types are not published or shipped. - Media and security instances are synthetic. No real companies, people or records.
- Downloads last month
- 30