AdithyaSK/data_agent

data_agent

Plain, Harbor-free version of the data-analysis agent tasks — usable directly via load_dataset. Splits: train 5000, test 250, eval 144. Deterministic grading, no LLM judge.

Columns

• —task_id, source_row_id — ids
• —question — the question to answer
• —answer — gold answer; reward_mode (numeric/exact_short/exact_bool/list/list_csv/flexible), atol/rtol — how to grade
• —difficulty_level (1-5), difficulty_tier (easy/medium/hard)
• —kaggle_dataset — source Kaggle dataset
• —hf_bucket, bucket_prefix — where the input files live on the HF Hub (fetch without Harbor)
• —files — input filenames; instruction — the full agent prompt
• —package_tier

Usage

from datasets import load_dataset
ds = load_dataset("AdithyaSK/data_agent", split="train")
row = ds[0]
print(row["question"], row["answer"], row["reward_mode"])

Getting the data files (no Harbor needed)

Files live in the HF bucket hf_bucket under bucket_prefix/:

from huggingface_hub import HfApi
api = HfApi()
api.snapshot_download(repo_id=row["hf_bucket"], repo_type="dataset",
                      allow_patterns=f"{row['bucket_prefix']}/*", local_dir="input")

Grading (deterministic, no LLM)

Use the bundled grader.py:

from grader import grade
r = grade(row["answer"], my_prediction, reward_mode=row["reward_mode"],
          abs_tol=row["atol"], rel_tol=row["rtol"])
print(r.reward)  # 1.0 if correct

Tiers: exact -> numeric(atol/rtol) -> list/percent normalization -> symbolic (math-verify).

Companion datasets

• —Harbor task suites (to run as environments via OpenEnv): AdithyaSK/data_agent_harbor_{train,test,eval}
• —SFT traces: AdithyaSK/data_agent_harbor_train_sft