microsoft/qlib · error · ValueError

Most of samples are dropped. Please check this task: {task}

Error message

Most of samples are dropped. Please check this task: {task}

What it means

When a MetaTaskDS (meta-learning dataset for data selection) processes a task in PROC_MODE_FULL, it drops NaN rows from the prepared train/test data. If the test segment loses more than 90% of its rows (or was empty to begin with), the task is judged unusable and this ValueError is raised, naming the offending task.

Source

Thrown at qlib/contrib/meta/data_selection/dataset.py:161

            please refer to the docs of _prepare_meta_ipt for detailed explanation.
        """
        super().__init__(task, meta_info)
        self.fill_method = fill_method

        time_perf = self._get_processed_meta_info()
        self.processed_meta_input = {"time_perf": time_perf}
        # FIXME: memory issue in this step
        if mode == MetaTask.PROC_MODE_FULL:
            # process metainfo_
            ds = self.get_dataset()

            # these three lines occupied 70% of the time of initializing MetaTaskDS
            d_train, d_test = ds.prepare(["train", "test"], col_set=["feature", "label"])
            prev_size = d_test.shape[0]
            d_train = d_train.dropna(axis=0)
            d_test = d_test.dropna(axis=0)
            if prev_size == 0 or d_test.shape[0] / prev_size <= 0.1:
                raise ValueError(f"Most of samples are dropped. Please check this task: {task}")

            assert (
                d_test.groupby("datetime", group_keys=False).size().shape[0] >= 5
            ), "In this segment, this trading dates is less than 5, you'd better check the data."

            sample_time_belong = np.zeros((d_train.shape[0], time_perf.shape[1]))
            for i, col in enumerate(time_perf.columns):
                # these two lines of code occupied 20% of the time of initializing MetaTaskDS
                slc = slice(*d_train.index.slice_locs(start=col[0], end=col[1]))
                sample_time_belong[slc, i] = 1.0

            # If you want that last month also belongs to the last time_perf
            # Assumptions: the latest data has similar performance like the last month
            sample_time_belong[sample_time_belong.sum(axis=1) != 1, -1] = 1.0

            self.processed_meta_input.update(
                dict(
                    X=d_train["feature"],

View on GitHub (pinned to 79633dd950)

Solutions

Shorten or shift the task's test segment so the label horizon (e.g. learned days) fits inside available data.
Check the underlying DataHandler output for the offending task: prepare the segment and inspect df.isna().mean() to see whether features or labels produce the NaNs.
Widen the data coverage (more dates/instruments) for the universe the task uses.
Catch this per-task and skip the bad task when building a MetaDatasetDS over many tasks.

Example fix

// before
task = {"dataset": {"kwargs": {"segments": {"test": ("2019-01-01", "2020-12-31")}}}}  # labels run past data end

// after
task = {"dataset": {"kwargs": {"segments": {"test": ("2019-01-01", "2020-06-30")}}}}  # label horizon fits

Defensive patterns

Strategy: try-catch

Validate before calling

d_test = task_handler.prepare("test", col_set=["feature", "label"])
prev = d_test.shape[0]
surv = d_test.dropna().shape[0]
if prev == 0 or surv / prev <= 0.1:
    logger.warning("task %s drops most samples; adjust segments", task)

Try / catch

good_tasks = []
for task in tasks:
    try:
        mtds = MetaTaskDS(task=task, ...)
        good_tasks.append(mtds)
    except ValueError as e:
        if "Most of samples are dropped" in str(e):
            logger.warning("skipping bad task: %s", task)
            continue
        raise

Prevention

Keep label horizons strictly inside available data for every segment.
Dry-run dropna() coverage checks per task before building the meta dataset.
Wrap per-task construction in try/except when iterating over many rolling tasks.

When it happens

Trigger: MetaTaskDS init or prepare where d_test.shape[0]/prev_size <= 0.1 after dropna — e.g. a task whose label window extends past available data, so almost every row has NaN labels.

Common situations: Task templates with test segments near the end of the available data (label horizon overruns the calendar); misconfigured segments in the task dict; instrument universes whose data ends before the segment; handlers producing mostly-NaN features.

Related errors

AI-assisted analysis of microsoft/qlib@79633dd950 (2026-08-15). Data as JSON: /api/errors/51af7ad87bd5f6f7. Report an issue: GitHub.