Datasets:

submissionpaper1234
/

mixture-then-select-selections

family stringclasses 4 values	method stringclasses 16 values	pool_index int64 0 50k	rank int64 0 999	role stringclasses 4 values	selection_id stringclasses 52 values	selection_seed int64 1 42 ⌀
reviewer_baseline	doremi_sft	36,624	0	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	8,182	1	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	8,889	2	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	9,447	3	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	38,230	4	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	29,473	5	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	7,499	6	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	8,646	7	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	33,568	8	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	17,249	9	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	43,944	10	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	6,597	11	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	6,270	12	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	2,389	13	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	25,999	14	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	24,718	15	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	44,674	16	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	20,509	17	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	13,017	18	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	9,696	19	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	41,056	20	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	17,346	21	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	26,150	22	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	9,560	23	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	34,972	24	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	19,705	25	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	48,000	26	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	6,668	27	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	28,373	28	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	26,337	29	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	18	30	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	31,928	31	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	19,286	32	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	8,375	33	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	22,835	34	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	888	35	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	38,870	36	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	31,658	37	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	16,044	38	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	16,453	39	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	11,750	40	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	10,743	41	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	46,909	42	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	42,405	43	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	33,882	44	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	15,474	45	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	29,183	46	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	2,148	47	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	4,792	48	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	8,869	49	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	14,958	50	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	47,331	51	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	40,093	52	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	9,419	53	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	43,085	54	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	26,390	55	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	27,097	56	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	8,323	57	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	4,588	58	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	29,207	59	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	40,149	60	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	10,149	61	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	22,902	62	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	42,360	63	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	22,477	64	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	9,473	65	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	11,736	66	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	5,818	67	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	41,250	68	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	46,028	69	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	36,572	70	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	17,819	71	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	13,869	72	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	8,015	73	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	28,334	74	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	2,423	75	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	10,919	76	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	9,946	77	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	10,326	78	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	49,120	79	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	41,395	80	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	29,329	81	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	43,551	82	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	42,049	83	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	13,110	84	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	24,638	85	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	49,243	86	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	34,756	87	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	48,457	88	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	24,100	89	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	4,835	90	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	17,062	91	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	37,017	92	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	21,978	93	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	6,877	94	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	18,206	95	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	868	96	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	8,679	97	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	32,807	98	baseline	doremi_sft	42
reviewer_baseline	doremi_sft	29,305	99	baseline	doremi_sft	42

End of preview. Expand in Data Studio

Frozen Qwen3-8B Selections

This dataset contains the exact selection metadata and pool indices used for a frozen cross-scale data-selection experiment. The subsets were selected using Qwen3-8B-derived information and can be transferred unchanged to a tokenizer-compatible larger target model.

The companion training and evaluation code is:

https://github.com/submissionpaper1234/mixture-then-select-reproducibility

Critical Interpretation

The selected instruction text was not generated by Qwen3-8B. Qwen3-8B was the selection-source model used to derive allocations or rankings over an existing instruction-tuning pool.

This repository publishes indices and metadata only. It does not redistribute the mixed-source instruction text or any model weights.

File	Purpose
`selection_bundle.json`	Canonical nested representation consumed by the training pipeline
`selection_index.jsonl`	Viewer-friendly table with one row per selection/index pair
`selection_manifest.json`	Bundle, table, pool, and per-selection checksums
`pool_manifest.json`	Exact canonical-pool contract and upstream source inventory
`materialize_subsets.py`	Validate a local canonical pool and export selected records
`validate_artifacts.py`	Standalone integrity and anonymity gate
`SOURCE_DATASETS.md`	Source and licensing notes

Coverage

The artifact contains 52 static subsets, each with exactly 1,000 unique indices:

Family	Subsets
Submitted MtS anchor and controls	3
Modern3 profiles and controls	35
TAP profiles	12
RegMix-SFT and DoReMi-SFT	2
Total	52

DIDS-SFT is intentionally absent from the static bundle. DIDS performs online, target-model-dependent sampling, so representing it as a frozen Qwen3-8B subset would change the baseline.

Canonical Pool Contract

Indices have meaning only for the exact ordered pool below:

Property	Value
Records	50,000
Bytes	65,881,018
SHA-256	`37155f4b639b29266e340f330d82408f6a9418489f7a275bad952569510de627`
Record format	JSON list; each record has formatted training text

The original build used 13 public instruction-tuning datasets, but exact upstream dataset revisions were not recorded. Therefore, rebuilding from the current upstream main revisions is not guaranteed to reproduce the canonical bytes. The provided tools fail closed unless the supplied pool matches the record count, byte size, and SHA-256.

Use With the Target-Scale Pipeline

Download this dataset:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="submissionpaper1234/mixture-then-select-selections",
    repo_type="dataset",
    local_dir="mixture-then-select-selections",
)

Copy or link selection_bundle.json into the code repository:

cp mixture-then-select-selections/selection_bundle.json \
  mixture-then-select-reproducibility/selections/selection_bundle.json

The code package validates the bundle hash, every index list, the target pool, and the complete 65-row matrix before launching.

Materialize One or More Static Subsets

To export records from an independently obtained canonical pool:

python3 materialize_subsets.py \
  --pool-file /absolute/path/to/train_pool.json \
  --selection-id submitted_mts \
  --selection-id regmix_sft \
  --output-dir ./materialized

To materialize all 52:

python3 materialize_subsets.py \
  --pool-file /absolute/path/to/train_pool.json \
  --all \
  --output-dir ./materialized

Materialized records preserve the selection-list order. Each export includes a manifest binding it to the canonical pool and selection hashes.

Frozen Transfer Rules

For a valid cross-scale comparison:

Freeze these static subsets before observing target-scale results.
Train every static method on its exact 1,000 indices.
Do not use the target model to rescore, rerank, recalibrate, or replace a frozen static subset.
Keep training and evaluation settings common across methods.
Run DIDS natively online and separately disclose its unique-example access, exposures, and compute.
Evaluate on IFEval, MMLU, GSM8K, and ARC-Challenge under the companion protocol.

Validation

python3 validate_artifacts.py

The validator checks:

The sealed bundle and table hashes.
Exactly 52 selections and 52,000 table rows.
Exactly 1,000 unique in-range indices per subset.
Every per-selection SHA-256.
Agreement between the nested bundle and tabular representation.
Absence of private filesystem provenance.

Limitations

Raw training text is not included because it combines upstream datasets with distinct licenses and terms.
Exact upstream revisions were not retained in the original pool build.
Static selections transfer Qwen3-8B selection information; they are not native target-model recalibrations.
Selection indices are unusable with a differently ordered or reformatted pool.
Task-conditioned Modern3 and TAP profiles should be interpreted using their declared objective names rather than as one universal selector.

Licensing

The selection indices, normalized metadata, manifests, and documentation are released under CC BY 4.0. The helper scripts are released under the MIT License. These licenses do not apply to upstream instruction datasets, models, or benchmarks. Obtain each upstream resource under its original terms.

Downloads last month: 6