name stringlengths 30 55 | instruction stringlengths 2.47k 7.16k | description stringclasses 1
value | keywords listlengths 3 3 | schema_version stringclasses 1
value | config unknown | metadata unknown | files listlengths 24 125 | location stringlengths 24 49 |
|---|---|---|---|---|---|---|---|---|
bespokelabs/act-tensor-sparse-panel-imputation-r2 | # Imputing a severely and heterogeneously incomplete firm-characteristic panel
Recover missing values in unseen firm characteristic panels as accurately as possible. Each panel
is an array with shape `(60 months, N firms, 45 characteristics)`.
Data:
* Every value has been cross-sectionally ranked within its `(month,... | [
"autoresearch",
"machine-learning",
"benchmark"
] | 1.4 | {
"schema_version": "1.4",
"artifacts": [
{
"source": "/var/lib/are",
"service": "grader"
}
],
"task": {
"name": "bespokelabs/act-tensor-sparse-panel-imputation-r2",
"version": "1.1.0",
"description": "",
"keywords": [
"autoresearch",
"machine-learning",
"be... | {
"category": "Machine Learning and AI",
"subcategory": "Unsupervised and representation learning",
"task_objective": [
"Implement",
"Optimize"
],
"artifact_type": [
"Single script or program"
],
"expert_time_estimate_hours": 4,
"are_version": "1.1",
"budget_seconds": 86400
} | [
"README.md",
"environment/.dockerignore",
"environment/Dockerfile",
"environment/app/data/dev/instance_00.npz",
"environment/app/data/dev/instance_01.npz",
"environment/app/data/dev/instance_02.npz",
"environment/app/data/dev/instance_03.npz",
"environment/app/data/dev/instance_04.npz",
"environment... | tasks/act-tensor-sparse-panel-imputation-r2 | |
bespokelabs/activeprune-al-unlabeled-pool-pruning | # Unlabeled-pool pruning for active learning
An active-learning loop trains a binary sentiment classifier over a large unlabeled pool of
documents. Each round, before the acquisition function runs, your code picks which subset of
the still-unlabeled documents the acquisition function is allowed to consider. Everything... | [
"autoresearch",
"machine-learning",
"benchmark"
] | 1.4 | {
"schema_version": "1.4",
"artifacts": [
{
"source": "/var/lib/are",
"service": "grader"
}
],
"task": {
"name": "bespokelabs/activeprune-al-unlabeled-pool-pruning",
"version": "1.1.0",
"description": "",
"keywords": [
"autoresearch",
"machine-learning",
"be... | {
"category": "Machine Learning and AI",
"subcategory": "Data Engineering and Curation",
"task_objective": [
"Implement",
"Optimize"
],
"artifact_type": [
"Single script or program"
],
"expert_time_estimate_hours": 4,
"are_version": "1.1",
"budget_seconds": 86400
} | [
"README.md",
"environment/.dockerignore",
"environment/Dockerfile",
"environment/app/data/dev_eval.jsonl.gz",
"environment/app/data/dev_eval_y.npy",
"environment/app/data/eval_a_X.npy",
"environment/app/data/eval_b_X.npy",
"environment/app/data/eval_c_X.npy",
"environment/app/data/pool_a.jsonl.gz",
... | tasks/activeprune-al-unlabeled-pool-pruning | |
bespokelabs/budgeted-covtype-dual-market-open | # Dual-market budgeted classification
Build a classification policy. It must learn while label prices move, and it must choose which
feature values to buy while feature prices move. The goal is to maximize balanced accuracy: higher
is always better, with no cutoff where further improvement stops counting. A constant-c... | [
"autoresearch",
"machine-learning",
"benchmark"
] | 1.4 | {
"schema_version": "1.4",
"artifacts": [
{
"source": "/var/lib/are",
"service": "grader"
}
],
"task": {
"name": "bespokelabs/budgeted-covtype-dual-market-open",
"version": "1.1.0",
"description": "",
"keywords": [
"autoresearch",
"machine-learning",
"benchm... | {
"difficulty": "hard",
"model_difficulty": "hard",
"category": "Machine Learning and AI",
"tags": [
"budgeted-acquisition",
"active-learning",
"dynamic-pricing",
"market-pricing"
],
"task_objective": [
"Optimize"
],
"artifact_type": [
"Generated output artifact"
],
"expert_t... | [
"README.md",
"environment/.dockerignore",
"environment/Dockerfile",
"environment/app/POLICY_TEMPLATE.py",
"environment/are-grader-ctx/Dockerfile",
"environment/are-grader-ctx/Dockerfile.are",
"environment/are-grader-ctx/are/drain.py",
"environment/are-grader-ctx/are/server.py",
"environment/are-grad... | tasks/budgeted-covtype-dual-market-open | |
bespokelabs/budgeted-imputation-mcar50 | # Recovering missing values under 50% missingness, on a fixed CPU budget
Fill in the missing values in unseen numeric tables as accurately as possible.
Each value is missing independently with 50% probability, and each call has a 30 second
wall clock limit.
Write an imputer at `/app/output/impute.py`.
## Submission ... | [
"autoresearch",
"machine-learning",
"benchmark"
] | 1.4 | {
"schema_version": "1.4",
"artifacts": [
{
"source": "/var/lib/are",
"service": "grader"
}
],
"task": {
"name": "bespokelabs/budgeted-imputation-mcar50",
"version": "1.1.0",
"description": "",
"keywords": [
"autoresearch",
"machine-learning",
"benchmark"
... | {
"category": "Machine Learning and AI",
"subcategory": "Feature engineering",
"task_objective": [
"Implement",
"Optimize"
],
"artifact_type": [
"Single script or program"
],
"expert_time_estimate_hours": 4,
"are_version": "1.1",
"budget_seconds": 86400
} | [
"README.md",
"environment/.dockerignore",
"environment/Dockerfile",
"environment/app/cellrunner.py",
"environment/app/cellspec.py",
"environment/app/data/dev/dev1.npz",
"environment/app/data/dev/dev2.npz",
"environment/app/data/dev/dev3.npz",
"environment/app/data/dev/dev4.npz",
"environment/app/d... | tasks/budgeted-imputation-mcar50 | |
bespokelabs/carps-star-discrepancy-subset-select | # Low-discrepancy subset selection
You get a cloud of `n` points in the unit cube `[0,1]^3` and a subset size `k`. Return the `k`
points that fill the cube most evenly, in the L-infinity star discrepancy sense.
For a set `P` of `k` points, the star discrepancy is:
$$d^*(P) = \sup_{q \in [0,1]^3} \left| \frac{\#\{p \... | [
"autoresearch",
"machine-learning",
"benchmark"
] | 1.4 | {
"schema_version": "1.4",
"artifacts": [
{
"source": "/var/lib/are",
"service": "grader"
}
],
"task": {
"name": "bespokelabs/carps-star-discrepancy-subset-select",
"version": "1.1.0",
"description": "",
"keywords": [
"autoresearch",
"machine-learning",
"ben... | {
"category": "Model Training and ML Infrastructure",
"subcategory": "Evaluation infrastructure",
"task_objective": [
"Implement",
"Optimize"
],
"artifact_type": [
"Single script or program"
],
"expert_time_estimate_hours": 3,
"are_version": "1.1",
"budget_seconds": 86400
} | [
"README.md",
"environment/.dockerignore",
"environment/Dockerfile",
"environment/app/data/examples/cloud_00.npy",
"environment/app/data/examples/cloud_01.npy",
"environment/app/data/examples/cloud_02.npy",
"environment/app/data/examples/cloud_03.npy",
"environment/app/data/examples/cloud_04.npy",
"e... | tasks/carps-star-discrepancy-subset-select | |
bespokelabs/causalpfn-cate-pehe-ihdp-surfaceb | # Heterogeneous treatment-effect estimation from observational data
Given observational data where treatment was not randomly assigned, estimate how much the
treatment would change the outcome for each held out row.
## Task
Write `/app/output/estimator.py` with a single entry point:
```python
def estimate_cate(X_tr... | [
"autoresearch",
"machine-learning",
"benchmark"
] | 1.4 | {
"schema_version": "1.4",
"artifacts": [
{
"source": "/var/lib/are",
"service": "grader"
}
],
"task": {
"name": "bespokelabs/causalpfn-cate-pehe-ihdp-surfaceb",
"version": "1.1.0",
"description": "",
"keywords": [
"autoresearch",
"machine-learning",
"benchm... | {
"category": "Machine Learning and AI",
"subcategory": "Model inference and prediction",
"task_objective": [
"Implement",
"Optimize"
],
"artifact_type": [
"Single script or program"
],
"expert_time_estimate_hours": 4,
"are_version": "1.1",
"budget_seconds": 86400
} | [
"README.md",
"environment/.dockerignore",
"environment/Dockerfile",
"environment/app/data/dev/realization_000.npz",
"environment/app/data/dev/realization_001.npz",
"environment/app/data/dev/realization_002.npz",
"environment/app/data/dev/realization_003.npz",
"environment/app/data/dev/realization_004.... | tasks/causalpfn-cate-pehe-ihdp-surfaceb | |
bespokelabs/causalrivers-heldout-station-graph-auroc | # Directed edge scores for river gauge subgraphs
A river gauge network is a directed graph over discharge measuring stations. An edge `i -> j`
means water measured at station `i` flows onward to station `j`, with no gauged station in
between. Given five years of discharge data from five stations at a time, score how l... | [
"autoresearch",
"machine-learning",
"benchmark"
] | 1.4 | {
"schema_version": "1.4",
"artifacts": [
{
"source": "/var/lib/are",
"service": "grader"
}
],
"task": {
"name": "bespokelabs/causalrivers-heldout-station-graph-auroc",
"version": "1.1.0",
"description": "",
"keywords": [
"autoresearch",
"machine-learning",
... | {
"category": "Machine Learning and AI",
"subcategory": "Model inference and prediction",
"task_objective": [
"Implement",
"Optimize"
],
"artifact_type": [
"Single script or program"
],
"expert_time_estimate_hours": 6,
"metric_name": "Mean per-subgraph AUROC of directed-edge scores over a se... | [
"README.md",
"environment/.dockerignore",
"environment/Dockerfile",
"environment/app/data/dev_a_edges.csv",
"environment/app/data/dev_b_edges.csv",
"environment/app/dev_data.py",
"environment/are-grader-ctx/Dockerfile",
"environment/are-grader-ctx/Dockerfile.are",
"environment/are-grader-ctx/are/dra... | tasks/causalrivers-heldout-station-graph-auroc | |
bespokelabs/coreset-selection-group-robust-waterbirds | # Coreset selection against an unlabeled slice
You're given embeddings for a pool of bird photos, produced once by a frozen image backbone.
Each embedding is 512-dimensional. Your job is to pick 240 of these rows to form the training
set for a fixed downstream classifier.
The classifier is scored on a held-out evalua... | [
"autoresearch",
"machine-learning",
"benchmark"
] | 1.4 | {
"schema_version": "1.4",
"artifacts": [
{
"source": "/var/lib/are",
"service": "grader"
}
],
"task": {
"name": "bespokelabs/coreset-selection-group-robust-waterbirds",
"version": "1.1.0",
"description": "",
"keywords": [
"autoresearch",
"machine-learning",
... | {
"category": "Model Training and ML Infrastructure",
"subcategory": "Data loading and training pipelines",
"task_objective": [
"Implement",
"Optimize"
],
"artifact_type": [
"Single script or program"
],
"expert_time_estimate_hours": 4,
"are_version": "1.1",
"budget_seconds": 86400
} | [
"README.md",
"environment/.dockerignore",
"environment/Dockerfile",
"environment/app/data/dev_draws.npy",
"environment/app/data/dev_pool_features.npy",
"environment/app/data/dev_pool_labels.npy",
"environment/are-grader-ctx/Dockerfile",
"environment/are-grader-ctx/Dockerfile.are",
"environment/are-g... | tasks/coreset-selection-group-robust-waterbirds | |
bespokelabs/cpu-decoder-graph-executor | # A faster executor for transformer-decoder compute graphs
`/app/reference_executor.py` runs transformer-decoder compute graphs on CPU. Write an
executor that computes the same thing faster.
## Submission format
- Deliverable: `/app/output/executor.py`, plus any helper modules placed beside it in
`/app/output/`.
-... | [
"autoresearch",
"machine-learning",
"benchmark"
] | 1.4 | {
"schema_version": "1.4",
"artifacts": [
{
"source": "/var/lib/are",
"service": "grader"
}
],
"task": {
"name": "bespokelabs/cpu-decoder-graph-executor",
"version": "1.1.0",
"description": "",
"keywords": [
"autoresearch",
"machine-learning",
"benchmark"
... | {
"category": "Model Training and ML Infrastructure",
"subcategory": "Systems and Efficiency",
"task_objective": [
"Optimize",
"Implement"
],
"artifact_type": [
"Single script or program",
"Codebase"
],
"expert_time_estimate_hours": 4,
"are_version": "1.1",
"budget_seconds": 86400
} | [
"README.md",
"environment/.dockerignore",
"environment/Dockerfile",
"environment/app/bench.py",
"environment/app/graph_spec.py",
"environment/app/harness.py",
"environment/app/reference_executor.py",
"environment/app/worker.py",
"environment/are-grader-ctx/Dockerfile",
"environment/are-grader-ctx/... | tasks/cpu-decoder-graph-executor | |
bespokelabs/cpu-llm-decode-throughput | # Serve a concurrent generation workload faster
Build a generation engine that serves the concurrent workload faster while returning the same
greedy token ids as the reference engine.
`/app/reference/reference_engine.py` serves generation requests one at a time on a 12-layer
causal language model (GPT-2 124M geometry... | [
"autoresearch",
"machine-learning",
"benchmark"
] | 1.4 | {
"schema_version": "1.4",
"artifacts": [
{
"source": "/var/lib/are",
"service": "grader"
}
],
"task": {
"name": "bespokelabs/cpu-llm-decode-throughput",
"version": "1.1.0",
"description": "",
"keywords": [
"autoresearch",
"machine-learning",
"benchmark"
... | {
"category": "Model Training and ML Infrastructure",
"subcategory": "Systems and Efficiency",
"task_objective": [
"Optimize",
"Implement"
],
"artifact_type": [
"Codebase",
"Single script or program"
],
"expert_time_estimate_hours": 6,
"are_version": "1.1",
"budget_seconds": 86400
} | [
"README.md",
"environment/.dockerignore",
"environment/Dockerfile",
"environment/app/data/dev_workload_0.json",
"environment/app/data/dev_workload_1.json",
"environment/app/data/dev_workload_2.json",
"environment/app/data/dev_workload_3.json",
"environment/app/data/warmup_workload.json",
"environmen... | tasks/cpu-llm-decode-throughput | |
bespokelabs/dctabeval-aeac-pooled-cat-statistics | # Access-request risk scoring from high-cardinality ID columns
Rank unseen access requests by how likely they are to be granted.
Employee access requests are recorded as nine nominal ID columns (resource, manager,
role rollups, department, title, family) and a binary outcome: the request was granted
or it was not. No... | [
"autoresearch",
"machine-learning",
"benchmark"
] | 1.4 | {
"schema_version": "1.4",
"artifacts": [
{
"source": "/var/lib/are",
"service": "grader"
}
],
"task": {
"name": "bespokelabs/dctabeval-aeac-pooled-cat-statistics",
"version": "1.1.0",
"description": "",
"keywords": [
"autoresearch",
"machine-learning",
"ben... | {
"category": "Machine Learning and AI",
"subcategory": "Feature engineering",
"task_objective": [
"Implement",
"Optimize"
],
"artifact_type": [
"Single script or program"
],
"expert_time_estimate_hours": 3,
"are_version": "1.1",
"budget_seconds": 86400
} | [
"README.md",
"environment/.dockerignore",
"environment/Dockerfile",
"environment/app/data/dev.csv",
"environment/app/local_panel.py",
"environment/are-grader-ctx/Dockerfile",
"environment/are-grader-ctx/Dockerfile.are",
"environment/are-grader-ctx/are/drain.py",
"environment/are-grader-ctx/are/serve... | tasks/dctabeval-aeac-pooled-cat-statistics | |
bespokelabs/fast-adv-budgeted-pgd50-robust-cifar10 | # Train an adversarially robust CIFAR 10 model
Train a `PreActResNet18` within 180 seconds. The grader attacks the trained model and measures how many unseen CIFAR 10 images remain correctly classified.
## Deliverable
Write `/app/output/train.py`. It must be a single Python file no larger than 102,400 bytes.
The gr... | [
"autoresearch",
"machine-learning",
"benchmark"
] | 1.4 | {
"schema_version": "1.4",
"artifacts": [
{
"source": "/var/lib/are",
"service": "grader"
}
],
"task": {
"name": "bespokelabs/fast-adv-budgeted-pgd50-robust-cifar10",
"version": "1.1.0",
"description": "",
"keywords": [
"autoresearch",
"machine-learning",
"b... | {
"category": "Model Training and ML Infrastructure",
"subcategory": "Training loops",
"task_objective": [
"Implement",
"Optimize",
"Secure or harden"
],
"artifact_type": [
"Single script or program",
"Generated output artifact"
],
"expert_time_estimate_hours": 4,
"metric_name": "PGD... | [
"README.md",
"environment/.dockerignore",
"environment/Dockerfile",
"environment/app/preact_resnet.py",
"environment/are-grader-ctx/Dockerfile",
"environment/are-grader-ctx/Dockerfile.are",
"environment/are-grader-ctx/are/drain.py",
"environment/are-grader-ctx/are/server.py",
"environment/are-grader... | tasks/fast-adv-budgeted-pgd50-robust-cifar10 | |
bespokelabs/fastercache-budgeted-video-dit-cache-policy | # Reduce video diffusion transformer evaluations
Video diffusion models are slow because they run a large transformer many times while refining
each video. CogVideoX 2b uses 50 diffusion steps and evaluates its transformer twice at each step,
once for the conditional prompt and once for the unconditional prompt. An un... | [
"autoresearch",
"machine-learning",
"benchmark"
] | 1.4 | {
"schema_version": "1.4",
"artifacts": [
{
"source": "/var/lib/are",
"service": "grader"
}
],
"task": {
"name": "bespokelabs/fastercache-budgeted-video-dit-cache-policy",
"version": "1.1.0",
"description": "",
"keywords": [
"autoresearch",
"machine-learning",
... | {
"category": "Machine Learning and AI",
"subcategory": "Model inference and prediction",
"task_objective": [
"Implement",
"Optimize"
],
"artifact_type": [
"Single script or program"
],
"expert_time_estimate_hours": 5,
"are_version": "1.1",
"budget_seconds": 86400
} | [
"README.md",
"environment/.dockerignore",
"environment/Dockerfile",
"environment/app/examples/example_policy.py",
"environment/app/harness/__init__.py",
"environment/app/harness/channel.py",
"environment/app/harness/hcore.py",
"environment/app/harness/run_eval.py",
"environment/app/harness/worker.py... | tasks/fastercache-budgeted-video-dit-cache-policy | |
bespokelabs/fastergcg-candidate-token-rank-ccc | # Rank token substitutions by true loss reduction
A prompt optimizer selects 64 possible token replacements using the gradient of its target loss. It would normally run the model once for every candidate to find which replacements help.
Build a ranker that predicts the same ordering without running the model at grade... | [
"autoresearch",
"machine-learning",
"benchmark"
] | 1.4 | {
"schema_version": "1.4",
"artifacts": [
{
"source": "/var/lib/are",
"service": "grader"
}
],
"task": {
"name": "bespokelabs/fastergcg-candidate-token-rank-ccc",
"version": "1.1.0",
"description": "",
"keywords": [
"autoresearch",
"machine-learning",
"bench... | {
"category": "Machine Learning and AI",
"subcategory": "Interpretability and model inspection",
"task_objective": [
"Implement",
"Optimize",
"Analyze"
],
"artifact_type": [
"Single script or program"
],
"expert_time_estimate_hours": 6,
"are_version": "1.1",
"budget_seconds": 86400
} | [
"README.md",
"environment/.dockerignore",
"environment/Dockerfile",
"environment/app/data/dev_behaviors.csv",
"environment/app/data/token_allowlist.npy",
"environment/app/example_ranker.py",
"environment/app/gcg_states.py",
"environment/are-grader-ctx/Dockerfile",
"environment/are-grader-ctx/Dockerf... | tasks/fastergcg-candidate-token-rank-ccc | |
bespokelabs/grpo-rl-halfcheetah-advantage-estimator | # Advantage estimation for a frozen PPO-clip harness
Design the advantage values used by a fixed PPO training program in Gymnasium's
`HalfCheetah-v4` environment. Your goal is to maximize the trained policy's average episode return
across hidden training and evaluation seeds.
The only part you control is the advantag... | [
"autoresearch",
"machine-learning",
"benchmark"
] | 1.4 | {
"schema_version": "1.4",
"artifacts": [
{
"source": "/var/lib/are",
"service": "grader"
}
],
"task": {
"name": "bespokelabs/grpo-rl-halfcheetah-advantage-estimator",
"version": "1.1.0",
"description": "",
"keywords": [
"autoresearch",
"machine-learning",
"... | {
"name": "advantage-estimation-half-cheetah",
"task_slug": "grpo-rl-halfcheetah-advantage-estimator",
"expert_time_estimate_hours": 4,
"title": "Advantage estimation for a frozen PPO-clip harness on HalfCheetah-v4 at a 1M-step budget",
"category": "Model Training and ML Infrastructure",
"subcategory": "Rei... | [
"README.md",
"environment/.dockerignore",
"environment/Dockerfile",
"environment/app/harness/estimator_worker.py",
"environment/app/harness/eval_policy.py",
"environment/app/harness/ppo_harness.py",
"environment/app/harness/train.py",
"environment/are-grader-ctx/Dockerfile",
"environment/are-grader-... | tasks/grpo-rl-halfcheetah-advantage-estimator | |
bespokelabs/hicard-latent-encoder | # Fixed-width encoding for a very high-cardinality categorical column
A tabular regression dataset has 10 numeric covariates `X` and one categorical column `G`.
`G` has `n_categories = 1600` levels with very different frequencies. Write an encoder that
replaces `G` with **at most 8 numeric columns** without ever seein... | [
"autoresearch",
"machine-learning",
"benchmark"
] | 1.4 | {
"schema_version": "1.4",
"artifacts": [
{
"source": "/var/lib/are",
"service": "grader"
}
],
"task": {
"name": "bespokelabs/hicard-latent-encoder",
"version": "1.1.0",
"description": "",
"keywords": [
"autoresearch",
"machine-learning",
"benchmark"
],
... | {
"category": "Machine Learning and AI",
"subcategory": "Feature engineering",
"task_objective": [
"Implement",
"Optimize"
],
"artifact_type": [
"Single script or program"
],
"expert_time_estimate_hours": 4,
"are_version": "1.1",
"budget_seconds": 86400
} | [
"README.md",
"environment/.dockerignore",
"environment/Dockerfile",
"environment/app/data/dev/instance_00000.npz",
"environment/app/data/dev/instance_00001.npz",
"environment/app/data/dev/instance_00002.npz",
"environment/app/data/dev/instance_00003.npz",
"environment/app/data/dev/instance_00004.npz",... | tasks/hicard-latent-encoder | |
bespokelabs/label-efficient-risk-estimator | # Label-efficient risk estimation
A classifier `f` assigns a probability distribution over `C` classes to each of `N = 2000`
unlabelled pool examples. Its risk on that pool is
$$
R = \frac{1}{N} \sum_i -\log \text{target\_probs}[i, y_i]
$$
where `y_i` is the true (unknown) class of example `i`. You may buy at most `... | [
"autoresearch",
"machine-learning",
"benchmark"
] | 1.4 | {
"schema_version": "1.4",
"artifacts": [
{
"source": "/var/lib/are",
"service": "grader"
}
],
"task": {
"name": "bespokelabs/label-efficient-risk-estimator",
"version": "1.1.0",
"description": "",
"keywords": [
"autoresearch",
"machine-learning",
"benchmark... | {
"category": "Machine Learning and AI",
"subcategory": "Evaluation, Calibration, and Robustness",
"task_objective": [
"Implement",
"Optimize"
],
"artifact_type": [
"Single script or program"
],
"expert_time_estimate_hours": 4,
"are_version": "1.1",
"budget_seconds": 86400
} | [
"README.md",
"environment/.dockerignore",
"environment/Dockerfile",
"environment/app/data/dev/dev00/labels.npy",
"environment/app/data/dev/dev00/meta.json",
"environment/app/data/dev/dev00/surrogate_probs.npy",
"environment/app/data/dev/dev00/target_probs.npy",
"environment/app/data/dev/dev01/labels.n... | tasks/label-efficient-risk-estimator | |
bespokelabs/less-is-more-pretrain-token-budget-selector | # Spend a pretraining token budget
Build a selector that chooses which web documents to use for language model pretraining under a fixed token budget.
## Goal
The document pool contains about 250 million GPT-2 BPE tokens. Your selector may spend 125,000,000 tokens.
The grader trains the same 124M parameter decoder-... | [
"autoresearch",
"machine-learning",
"benchmark"
] | 1.4 | {
"schema_version": "1.4",
"artifacts": [
{
"source": "/var/lib/are",
"service": "grader"
}
],
"task": {
"name": "bespokelabs/less-is-more-pretrain-token-budget-selector",
"version": "1.1.0",
"description": "",
"keywords": [
"autoresearch",
"machine-learning",
... | {
"task_slug": "less-is-more-pretrain-token-budget-selector",
"title": "Spend a pretraining token budget: write the corpus selector that minimises held-out perplexity",
"category": "Model Training and ML Infrastructure",
"subcategory": "Data loading and training pipelines",
"expert_time_estimate_hours": 8,
... | [
"README.md",
"environment/.dockerignore",
"environment/Dockerfile",
"environment/app/harness/dev_cycle.py",
"environment/app/harness/pretrain_harness.py",
"environment/app/harness/train_config.json",
"environment/app/output/select.py",
"environment/are-grader-ctx/Dockerfile",
"environment/are-grader... | tasks/less-is-more-pretrain-token-budget-selector | |
bespokelabs/mlh-coco-16bit-hash-head-map5000 | # 16-bit hash codes for multi-label image retrieval
Build a binary encoder that maps a frozen 512-dimensional image feature vector to a 16-bit code,
so that the Hamming distance between two codes reflects how much semantic content the two images
share.
## Data
- `/app/data/train_features.npy`: shape `(10000, 512)`, ... | [
"autoresearch",
"machine-learning",
"benchmark"
] | 1.4 | {
"schema_version": "1.4",
"artifacts": [
{
"source": "/var/lib/are",
"service": "grader"
}
],
"task": {
"name": "bespokelabs/mlh-coco-16bit-hash-head-map5000",
"version": "1.1.0",
"description": "",
"keywords": [
"autoresearch",
"machine-learning",
"benchma... | {
"category": "Machine Learning and AI",
"subcategory": "Unsupervised and representation learning",
"task_objective": [
"Implement",
"Optimize"
],
"artifact_type": [
"Single script or program"
],
"expert_time_estimate_hours": 4,
"are_version": "1.1",
"budget_seconds": 86400
} | [
"README.md",
"environment/.dockerignore",
"environment/Dockerfile",
"environment/app/data/train_features.npy",
"environment/app/data/train_labels.npy",
"environment/are-grader-ctx/Dockerfile",
"environment/are-grader-ctx/Dockerfile.are",
"environment/are-grader-ctx/are/drain.py",
"environment/are-gr... | tasks/mlh-coco-16bit-hash-head-map5000 | |
bespokelabs/reppo-reliable-onpolicy-control-trainer | # Train control policies that hit a performance bar on every seed and hold it
Write a reinforcement-learning training procedure at `/app/solution/trainer.py`. The score is
the **reliable-success fraction**. It is the fraction of runs whose policy reaches 90% of a
reference level and stays there for the last ten evalua... | [
"autoresearch",
"machine-learning",
"benchmark"
] | 1.4 | {
"schema_version": "1.4",
"artifacts": [
{
"source": "/var/lib/are",
"service": "grader"
}
],
"task": {
"name": "bespokelabs/reppo-reliable-onpolicy-control-trainer",
"version": "1.1.0",
"description": "",
"keywords": [
"autoresearch",
"machine-learning",
"... | {
"task_slug": "reppo-reliable-onpolicy-control-trainer",
"category": "Model Training and ML Infrastructure",
"subcategory": "Model Training and Alignment",
"metric_name": "reliable_success_fraction",
"objective": "maximize",
"graded_deliverable": "/app/solution/trainer.py",
"device": "cpu",
"expert_tim... | [
"README.md",
"environment/.dockerignore",
"environment/Dockerfile",
"environment/app/child_main.py",
"environment/app/envs.py",
"environment/app/panel.py",
"environment/app/public_panel.json",
"environment/app/run_public.py",
"environment/app/solution/trainer.py",
"environment/are-grader-ctx/Docke... | tasks/reppo-reliable-onpolicy-control-trainer | |
bespokelabs/sae-sparse-dict-nmse-frontier | # Sparse coding of GPT-2 activations
Build a sparse dictionary that reconstructs GPT-2 small layer 8 residual stream activations.
## Goal
The grader measures normalized reconstruction error on 1,048,576 hidden activations:
$$
\frac{\sum_i \lVert x_i-\hat{x}_i\rVert^2}
{\sum_i \lVert x_i-\bar{x}\rVert^2}.
$$
Lower ... | [
"autoresearch",
"machine-learning",
"benchmark"
] | 1.4 | {
"schema_version": "1.4",
"artifacts": [
{
"source": "/var/lib/are",
"service": "grader"
}
],
"task": {
"name": "bespokelabs/sae-sparse-dict-nmse-frontier",
"version": "1.1.0",
"description": "",
"keywords": [
"autoresearch",
"machine-learning",
"benchmark"... | {
"category": "Machine Learning and AI",
"subcategory": "Unsupervised and representation learning",
"task_objective": [
"Optimize",
"Implement"
],
"artifact_type": [
"Single script or program",
"Generated output artifact"
],
"expert_time_estimate_hours": 6,
"are_version": "1.1",
"budge... | [
"README.md",
"environment/.dockerignore",
"environment/Dockerfile",
"environment/app/extract_acts.py",
"environment/are-grader-ctx/Dockerfile",
"environment/are-grader-ctx/Dockerfile.are",
"environment/are-grader-ctx/are/drain.py",
"environment/are-grader-ctx/are/server.py",
"environment/are-grader-... | tasks/sae-sparse-dict-nmse-frontier | |
bespokelabs/shortest-valid-ci-l2-ece | # Shortest valid confidence interval for the squared l2 calibration error
Given predicted class probabilities and their true labels, return a confidence interval for the
squared calibration error defined below. The interval should be as short as possible while still
covering the true value in at least 88% of repeated ... | [
"autoresearch",
"machine-learning",
"benchmark"
] | 1.4 | {
"schema_version": "1.4",
"artifacts": [
{
"source": "/var/lib/are",
"service": "grader"
}
],
"task": {
"name": "bespokelabs/shortest-valid-ci-l2-ece",
"version": "1.1.0",
"description": "",
"keywords": [
"autoresearch",
"machine-learning",
"benchmark"
... | {
"category": "Machine Learning and AI",
"subcategory": "Model evaluation and benchmarking",
"task_objective": [
"Implement",
"Optimize"
],
"artifact_type": [
"Single script or program"
],
"expert_time_estimate_hours": 6,
"are_version": "1.1",
"budget_seconds": 86400
} | [
"README.md",
"environment/.dockerignore",
"environment/Dockerfile",
"environment/app/data/dev_panel.json",
"environment/app/panel.py",
"environment/are-grader-ctx/Dockerfile",
"environment/are-grader-ctx/Dockerfile.are",
"environment/are-grader-ctx/are/drain.py",
"environment/are-grader-ctx/are/serv... | tasks/shortest-valid-ci-l2-ece | |
bespokelabs/sketched-newton-cov-estimator | # Limiting covariance of an online sketched Newton method
Estimate the optimizer's limiting covariance matrix $\Xi$ from the records of one run.
The fixed optimizer in `/app/harness/runner.py` processes one linear regression sample at a time.
Its final estimate varies across repeated runs. After scaling by the final ... | [
"autoresearch",
"machine-learning",
"benchmark"
] | 1.4 | {
"schema_version": "1.4",
"artifacts": [
{
"source": "/var/lib/are",
"service": "grader"
}
],
"task": {
"name": "bespokelabs/sketched-newton-cov-estimator",
"version": "1.1.0",
"description": "",
"keywords": [
"autoresearch",
"machine-learning",
"benchmark"... | {
"category": "Model Training and ML Infrastructure",
"subcategory": "Evaluation infrastructure",
"task_objective": [
"Implement",
"Optimize"
],
"artifact_type": [
"Single script or program"
],
"expert_time_estimate_hours": 6,
"are_version": "1.1",
"budget_seconds": 86400
} | [
"README.md",
"environment/.dockerignore",
"environment/Dockerfile",
"environment/app/harness/dev/inst_0.json",
"environment/app/harness/dev/inst_0_xi_star.npy",
"environment/app/harness/dev/inst_1.json",
"environment/app/harness/dev/inst_1_xi_star.npy",
"environment/app/harness/runner.py",
"environm... | tasks/sketched-newton-cov-estimator |
End of preview. Expand in Data Studio
AutoResearchExam
This repository contains the 29 tasks from AutoResearchExam. Version 1.1 runs natively on Harbor with any agent harness. Leaderboard scores come from the official harness on version 1.0. See Run for both.
News
- 2026-10: v1.1. Run AutoResearchExam with any agent harness using plain
harbor runon Harbor main, without depending on any custom harness. Same tasks, data and graders as v1.0. Leaderboard scores still come from v1.0 with the official harness. - 2026-09: v1.0. Initial release with the official harness and leaderboard.
Tasks by area
View tasks by area →
Model training
Representation learning
- Compact encoding for a column with many categories
- Compact item embeddings for recommendations
- Short binary codes for image search
Reinforcement learning
Predictive modeling
Model merging and distillation
Algorithms and optimization
Statistical methods
- Estimate uncertainty in a streaming optimizer
- Estimate individual treatment effects
- Recover river flow links
Constrained optimization
Data engineering and curation
Imputation
Data selection
- Unlabeled pool pruning
- Low discrepancy subset selection
- Group robust coreset selection
- Pretraining token budget selection
Systems and efficiency
Inference
Compression
Evaluation, calibration, and robustness
Metric estimation
Robustness
AI safety and alignment
Interpretability
Run
Each run has a 24 hour research budget, and the agent should keep working until it ends.
With the official harness
The leaderboard uses the v1.0 tasks with the official AutoResearchExam-Terminus harness. To reproduce leaderboard scores, follow the instructions in that repository.
With your own harness
The official harness keeps the agent working and submitting until the budget ends. With your own harness, that is up to the agent. An agent can stop early, but then it scores lower, because it did not follow the instruction to keep going.
Install Harbor, start Docker, and set up your model provider's credentials. Then run any agent harness on a task:
git clone https://github.com/bespokelabsai/AutoResearchExam.git
harbor run \
--path AutoResearchExam/cpu-decoder-graph-executor \
--agent <agent> --model <model> --env docker \
--ak prompt_template_path=AutoResearchExam/are_protocol.md.j2
Use --path AutoResearchExam to run all tasks. are_protocol.md.j2 adds the research protocol
after each task instruction: submit regularly, keep improving the solution, and keep going
until the run ends.
During the run, the agent can reach api.openai.com and api.anthropic.com. If your harness
uses another host, add --allow-agent-host <host>.
GPU tasks need an NVIDIA GPU and the NVIDIA Container Toolkit. Grading shares GPU 0 with the
agent by default. On a machine with a spare GPU, set ARE_GRADER_GPU=1 to give grading its
own GPU.
How autoresearch submissions and intermediate grading work
Each task runs the agent next to two grader containers.
- The agent submits with the
submittool of theareMCP server, or over HTTP athttp://grader:8000/submit. The first grader scores the submission on the public validation split. The agent gets the score and grader output from theresulttool. - The second grader scores the same submission on the private test split. The agent cannot reach it and never sees these scores.
- Grading does not pause the agent. The agent keeps working while a submission is graded. Only one grade runs at a time, validation or test. A new submission is refused until the previous one has its validation score.
- When the agent stops, the final workspace is graded too, unless it matches the last submission.
How scores can differ from v1.0
In v1.0, the official harness keeps the agent submitting for the full 24 hours. In v1.1, that is your agent's responsibility.
In v1.0, grading paused the agent. Now the agent keeps working while a submission is graded, so it has more working time. Because only one submission is graded at a time, an agent cannot flood the grader.
Grading runs on the same machine as the agent and can compete with it for compute. A grader uses as many cores as the task gives the agent, and the two graders never run at the same time, so plan for about twice the cores each task requests.
Scores
Each trial writes verifier/reward.json. Its reward is the private test score of the
submission with the best validation score. Its auarc uses each task's own score scale. To
put scores on the leaderboard scale, follow the
scoring guide
to normalize scores and compute AUARC. Every submission's scores and times are in the
trial's artifacts/var/lib/are/ledger.jsonl.
Citation
If you use AutoResearchExam in your research, please cite:
@misc{ramesh2026autoresearchexam,
author = {Ramesh, Anirudha and
Devic, Siddartha and
Garg, Shivank and
Parulekar, Advait and
Mahnot, Drish and
Pimpalgaonkar, Shreyas and
Suresh, Vishnu and
Dimakis, Alex and
Sathiamoorthy, Maheswaran},
title = {{AutoResearchExam}},
year = {2026},
month = sep,
url = {https://benchmarks.bespokelabs.ai/autoresearchexam/},
note = {Published September 9, 2026}
}
- Downloads last month
- 695