jax-tasks-v1

Task dataset for a JAX RL / eval environment, in the shape used by the Prime Intellect Environments Hub.

38 JAX tasks across 5 categories. Each task gives the model one or more input arrays and an instruction; the answer is the array left in result, graded with numpy.allclose against a reference. Grading is deterministic — no LLM judge, no external API, CPU only.

Category	Tasks	Covers
array_ops	14	reshape, transpose, axis reductions, clip, sort/argsort, cumsum, matmul, row-normalise, jnp.where, functional index update (.at[].set() / .at[].add()), diff
nn	8	softmax, log-softmax, relu, gelu, sigmoid, an explicit dense layer, MSE, per-row cross-entropy
autodiff	7	jax.grad, gradient through sin, elementwise second derivative, jax.vmap, jax.jacfwd, jax.value_and_grad, jax.jit
linalg	6	symmetric eigenvalues, Cholesky, solve, SVD, inverse, trace
random	3	PRNGKey, split, normal, uniform, permutation — all with the seed pinned in the prompt

Fields

Field	Description
task_id	stable id, e.g. jax-017
category	one of the five above
prompt	the natural-language instruction shown to the model
input_data	JSON object mapping array name (x, m, b, s, l) to a serialised array
expected_output	the serialised reference result

How it was built, and why you can trust the answer key

Tasks are defined as (deterministic input arrays, instruction, reference solution). The expected output is computed by executing the reference, never written by hand.

Every task is then independently verified: the reference runs and returns a real-valued array; it is deterministic across two executions; the result is finite, non-empty, and not identical to its input; and both the result and every input survive the serialisation round-trip exactly, dtype and shape included. All 38 pass on jax 0.11.0.

That verifier earned its place here: it caught two real defects on the first run of this dataset — a "sort ascending" task whose input was already a linspace and therefore an identity task, and a second-derivative task written as grad(grad(f)), which JAX rejects because the inner gradient of a sum over a vector is itself a vector and grad requires scalar output.

Two JAX-specific decisions

1. float32 throughout; jax_enable_x64 deliberately left off. JAX defaults to 32-bit and silently downcasts. Building the key under x64 and grading under the default would make every float task fail for a reason unrelated to the model.
2. PRNG keys are explicit (jax.random.PRNGKey(0)), and the seed is stated in the prompt. JAX's threefry stream is reproducible, which is exactly what makes these gradeable — but only if the seed is pinned rather than assumed.

Builder and verifier: build_tasks.py.