jayden8888/acorn-streams

Acorn Streams v0.2

中文说明

Acorn Streams is a stream-native dataset for studying which new experiences an adaptive language model should consolidate into slow weights and which it should learn temporarily. The data unit is an ordered stream. Each episode contains an adaptation support set, held-out queries, unlabeled consolidation probes, a closed candidate set, revision metadata, and hidden future-demand labels used only for evaluation.

Version 0.2 is a data-quality release. It increases genuinely independent support examples, makes the order intervention strictly paired, and adds a conservatively filtered natural few-shot task-induction configuration. CurLL and unrelated external benchmarks from v0.1 are intentionally not part of this training-focused snapshot.

Size and why it is not an arbitrary 50k

The release contains 4,533 streams, 36,264 episode rows, and 4,614,336 nested support/query/probe items.

Configuration	Train	Validation	Test	Total episodes	Independent content
controlled	24,000	4,800	7,200	36,000	1,500 seed groups / 12,000 unique episodes
adaptable	184	40	40	264	165 source tasks

The 36,000 controlled rows are 1,500 independent content seed groups × 3 exact order-only permutations × 8 episodes. Each unique controlled episode contains 32 support, 64 query, and 32 probe items. Consequently, the release already has about 1.54 million unique controlled items before order replication. Expanding to 50k by duplicating orders or templates would inflate the row count without adding learning signal. For training, use one order per seed group or sample an order per epoch; reserve the other paired orders for order-sensitivity analysis.

Configurations

controlled

The primary causal benchmark contains hidden symbolic mappings, latent linear rules, fresh DSL/API semantics, key-value bindings, revisions, recurrences, transient tasks, and noise. Support size is 32 genuinely distinct task inputs—not one small set repeated under cosmetic prompt wrappers.

Train/validation/test use disjoint seed groups: 1,000 / 200 / 300. For every seed, orders 0, 1, and 2 contain exactly the same rules and support/query/probe examples. Only the valid topological interleaving changes, and within-task recurrence/revision chronology is preserved.

adaptable

This auxiliary natural configuration derives from the Apache-2.0 UnpredicTable data used by AdapTable. A web table is treated as a task and its rows as demonstrations. All 30 clustered upstream configurations were scanned. The released 165 tasks pass task-level filters for:

• at least 24 conflict-free rows;
• 2–8 repeated labels with bounded target length;
• no identical input/output pair;
• no detected URL, email, phone-like string, or risky free-text output column;
• target-in-input leakage at or below 5%;
• exact task deduplication and source-task-disjoint splits.

Every episode uses an 8/8/8 support/query/probe split. Support covers every candidate label, and query/probe source rows are disjoint from support. This configuration is best used for auxiliary transfer and natural-task validation, not as a replacement for the controlled causal benchmark.

Loading

from datasets import load_dataset

controlled = load_dataset("jayden8888/acorn-streams", "controlled")
natural = load_dataset("jayden8888/acorn-streams", "adaptable")

row = controlled["train"][0]
print(row["stream_id"], row["episode_index"])
print(row["support"][0]["prompt"], row["support"][0]["target"])

Reconstruct complete streams by grouping on stream_id and sorting by episode_index. Acorn code can load the same standard Parquet layout directly:

from acorn.data.hf_dataset import PublishedDatasetConfig, load_published_streams

streams = load_published_streams(
    PublishedDatasetConfig(
        repo_id="jayden8888/acorn-streams",
        configuration="controlled",
        split="train",
    )
)

Row schema

Field	Meaning
dataset_version, configuration, split, source	Release and provenance.
stream_id, stream_seed, stream_order, seed_group	Stream identity and pairing coordinates.
episode_index, episode_id, task_id, task_family	Episode position and task identity.
experience_type, slot_role	Evaluation labels; never expose them to the model.
support	Examples available to fast adaptation.
query	Held-out closed-set evaluation examples; never train on them.
probe_prompts	Unlabeled, query-disjoint consolidation prompts.
future_reuse_prob, n_future_occurrences	Privileged future-demand labels.
version, revises	Revision-chain metadata.
*_metadata_json	Lossless residual metadata.

Each support/query item contains prompt, target, candidates, optional source and source_id, and metadata_json.

Correct training use

• Adapt only on support.
• Evaluate only on query.
• Use probe_prompts without targets for consolidation.
• Never feed experience_type, slot_role, future_reuse_prob, n_future_occurrences, future episodes, or query targets into the model/controller.
• Select hyperparameters on validation only and freeze them before test.
• Report results by independent seed group, not by treating three paired orders as three independent samples.

Quality evidence

Every stream was reconstructed through Acorn's native dataclasses. The automated release gate verifies schema invariants, closed candidates, support/query/probe disjointness, latent-label leakage, split isolation, and seed isolation. All three controlled splits achieved 100% content-identical order pairing with valid within-task chronology. Natural splits have zero exact query-prompt overlap.

Machine-readable evidence is included in:

• stats/build_summary.json
• stats/quality_report.json
• stats/selection_report.json
• dataset_info.json (row counts, schema, byte sizes, and SHA-256 checksums)

Limitations

The controlled families are procedurally generated abstractions, so conclusions must still be validated on natural tasks. UnpredicTable originates from web tables and can retain semantic ambiguity despite conservative automated filtering; it is therefore an auxiliary set. The natural subset is intentionally small because task quality and independence were prioritized over row volume.

Licenses and citation

• Controlled configuration and Acorn dataset tooling: MIT.
• adaptable: derived from UnpredicTable / AdapTable, Apache-2.0; see licenses/ADAPTABLE-LICENSE-NOTE.md.

If you use adaptable, cite both Acorn Streams and Few-shot Adaptation Works with UnpredicTable Data (ACL 2023). See CITATION.cff.