eric-ml-nlp/Delveta-LayaChoice-v2-Data

Delveta LayaChoice-v2 Data

1. Dataset Overview

LayaChoice-v2 is a multiple-choice dataset for capability selection: given a user request and a set of candidate capability cards, choose the card that should handle the request, or choose the single REJECT card when no candidate is correct.

The dataset is built for fine-tuning and evaluating a Laya-style non-autoregressive decision model. Each question contains:

• a user request (input.query);
• K real recalled capability cards, where K is 0 to 3;
• exactly one frozen REJECT card.

Therefore every row has K + 1 options (1 to 4). The option count is dynamic; the fixed constant OPTION_SLOTS = 4 is only the upper bound used by the training adapter. REJECT is a real option, not a threshold or a post-hoc rule.

The default configuration is recall_threshold_082, the bundle used for formal training. Two historical thresholds, recall_threshold_085 and recall_threshold_088, are archived under data_versions/. All three versions share the same source rows, the same split assignment and one shared recall pass; they differ in the similarity threshold that decides which capabilities become candidates.

Task	Pick one of K capability cards plus one REJECT card
Rows per version	48,333 (train 36,197 / validation 2,407 / test 9,729)
Versions	threshold 0.82 (default), 0.85, 0.88
Capabilities	18
Languages	en 22,053 / zh 20,667 / mixed 5,613
Option count	1 to 4 (K + 1)
Training rows	36,197, of which 400 are held out for calibration
Effective gradient rows	35,797
Card view	B_noprov (production card without query examples and provenance)

2. Dataset Structure

Delveta-LayaChoice-v2-Data/
├── README.md
├── DATASET_SCHEMA.md
├── SHA256SUMS
├── train.jsonl                  # default: recall_threshold_082
├── val.jsonl
├── test.jsonl
├── manifest.json
└── data_versions/
    ├── README.md
    ├── recall_threshold_085/
    │   ├── train.jsonl
    │   ├── val.jsonl
    │   ├── test.jsonl
    │   ├── manifest.json
    │   └── SHA256SUMS
    └── recall_threshold_088/
        ├── train.jsonl
        ├── val.jsonl
        ├── test.jsonl
        ├── manifest.json
        └── SHA256SUMS

The root files are the default recall_threshold_082 bundle. Each historical directory is an independent, self-contained bundle with its own JSONL files, manifest and checksums. See DATASET_SCHEMA.md for the complete field contract and validation rules, and data_versions/README.md for provenance and version verification.

The JSONL split file names are train.jsonl, val.jsonl and test.jsonl. In the Hugging Face configuration above, val.jsonl is exposed as the validation split.

Bundle sizes

Sizes are computed from the shipped bytes in decimal units (1 MB = 1,000,000 bytes).

Version	train.jsonl	val.jsonl	test.jsonl	manifest.json	Total
0.82 (root)	69.01 MB	4.64 MB	18.64 MB	0.71 MB	93.00 MB
0.85	54.63 MB	3.62 MB	14.73 MB	0.71 MB	73.69 MB
0.88	47.45 MB	3.16 MB	12.81 MB	0.71 MB	64.13 MB

3. Dataset Schema

Every row is a four-key record. Only input (query + options) and label are training fields; message_id and language are metadata. The split is not stored in the row; the containing file name is the authority.

Key	Type	Present in every row	Used to build model input / target	Meaning
message_id	string	yes	no — metadata only	Canonical uuid4 sample id, unique within the bundle
language	string	yes	no — metadata only	Resolved language: en, zh or mixed
input	object	yes	yes — model input	Model input container; exactly query and options
label	integer	yes	yes — scalar class index	0-based index of the correct option in input.options; the selected option maps to a capability ID or REJECT.

All four keys are present in every row, but only input and label participate in training. message_id and language are metadata: they never enter the model input, the target or calibration.

input has exactly two keys, both used by the model:

Key	Type	Present in every row	Used to build model input / target	Meaning
input.query	string	yes	yes — request segment	The user request shown to the model
input.options	list[string]	yes	yes — one [MASK] slot each	The exact option texts, in stored option order

Each option string is "<capability_id>: <card>", except that an empty card is the bare capability id. The REJECT option is "REJECT: <frozen reject card>".

Two invariants hold for every row:

1. input.options has 1 to 4 entries and contains exactly one option whose id is REJECT.
2. 0 <= label < len(input.options).

For a capability row, the option at label is a capability card. For a REJECT row, the option at label is the REJECT card. The stored option order is the shuffled order; it does not encode retrieval rank.

message_id is a fresh uuid4 in every threshold bundle and is unique within a bundle. It is sample metadata: it never enters the model input, the target or calibration.

3.1 Adapter view

The training adapter derives three values from input:

Derived value	Source
order	capability id of each option, in stored option order
options	card text of each option (the part after "<id>: ")
gold_index	label

target_kind, gold_capability_id and reject_index are derived properties of those values. They are not stored in a row.

3.2 Fields used by training

Row field	Use
input.query	Appended to the model input as the request segment
input.options	Each option is prefixed with one [MASK] marker and placed in the head
label	Scalar class index over the row's actual option slots; expanded at training time to a one-hot vector (length K+1)
message_id	Metadata only; not model input, not target
language	Metadata only; not model input, not target

The instruction is not a row field. It comes from manifest.instructions (Which capability should handle the user's request?), read at render time.

3.3 Final model input

The renderer builds one flat token sequence per row:

[CLS] choice question: <instructions> [SEP]
        [MASK] <option 0> [MASK] <option 1> ... [SEP]
        <query> [SEP]

• [CLS], [SEP] and [MASK] are the base model's special tokens.
• input.options are emitted in their stored order, one [MASK] marker per option.
• input.query is appended after the option block.
• label never appears in the sequence. It is stored as an integer slot index and expanded at training time into a one-hot vector over the row's actual option slots (length K+1, 1–4). This is not a global 0–17 capability-class one-hot vector.

For the row in §6.7 (K = 2, label = 0), the sequence is:

[CLS] choice question: Which capability should handle the user's request? [SEP]
[MASK] cap-translate: ### cap-translate … [MASK] cap-pdf-extract-text: ### cap-pdf-extract-text … [MASK] REJECT: ### REJECT … [SEP]
Can you say “Labels must be supported by the original sentence. Please confirm whether the final output preserves every citation.” in Chinese? [SEP]

The … shortens the card bodies; the stored option strings are used verbatim. label = 0 selects the first [MASK] slot.

Budget: each option's text is capped at 256 tokens (the [MASK] marker is additional), the question head and all options share 768 tokens, and the whole sequence is capped at 1024 tokens. No padding is added when the sequence is shorter than the cap.

4. Dataset Splits

Split	Rows	Share	en	zh	mixed	Type
train	36,197	74.89%	16,499	15,505	4,193	gradient training; 400 held out for calibration
val	2,407	4.98%	1,116	1,030	261	checkpoint selection
test	9,729	20.13%	4,438	4,132	1,159	final benchmark only
total	48,333	100%	22,053	20,667	5,613

Splits are stratified by source gold label × resolved language and allocated with family-aware indivisible units. The target is about 75/20/5, with minimum representation for small groups:

• group size 1: all rows in train;
• group size 2: one train, one test;
• group size 3 or more: at least one validation and one test row, converging to 75/20/5 for large groups.

Source gold label and resolved language define the strata; rows from the same query family never cross splits. One split assignment is computed once and reused by all three thresholds.

Calibration. The first 400 rows selected from the final post-shuffle train order (seed 20260922, rule min(400, n_train // 10)) form the calibration split. Those rows remain in train.jsonl but are excluded from gradient updates, leaving 35,797 effective training rows. manifest.json records the indices; membership is positional, reproducible and disjoint from the effective train set. Calibration is used for temperature fitting, not for checkpoint selection.

The test split is never used for training or checkpoint selection. Checkpoint selection uses validation top-1 only.

Option-count distribution

Version	K=0	K=1	K=2	K=3	Mean K
0.82	1,056	18,000	14,955	14,322	1.880206
0.85	1,938	32,460	10,665	3,270	1.315871
0.88	2,664	41,420	4,036	213	1.037200

K counts capability cards only. K=0 rows contain the REJECT card alone (one option).

5. Dataset Versions

All versions contain the same 48,333 source queries in the same positional order in each split, the same split assignment and the same 400 calibration indices. message_id is a fresh uuid4 per bundle, so it is not shared across versions and is not an alignment key. Candidate sets, labels and therefore option counts differ only because the similarity threshold differs.

Version	Threshold	Gold = capability	Gold = REJECT	K=0 / 1 / 2 / 3	Repaired K=0→K=1
recall_threshold_082 (default)	0.82	44,201	4,132	1,056 / 18,000 / 14,955 / 14,322	152
recall_threshold_085	0.85	44,009	4,324	1,938 / 32,460 / 10,665 / 3,270	603
recall_threshold_088	0.88	44,061	4,272	2,664 / 41,420 / 4,036 / 213	2,466

A lower threshold admits more candidates; a higher threshold admits fewer and produces more REJECT labels. The differences are structural, not noise: the three bundles are not resampled or relabeled copies of one another, and their training rows differ.

Frozen checksums

Each bundle has a SHA256SUMS file covering its manifest and three JSONL files. The root SHA256SUMS additionally covers this README and DATASET_SCHEMA.md.

File	recall_threshold_082 (default, repository root)	recall_threshold_085	recall_threshold_088
manifest.json	92514b46c1e090d50bcae9464112aa80ad8ba48f4e929f098afb929d7488adf8	7443cc774fce07319068497848f762cb5639cc5e15e98ffe1e4a8b7a4b95ea70	dd12efeb315c1bcfd9a7a1ba1aeafb1ced57416e6c01e40ee3df85d9169a5395
train.jsonl	97ed1b54a205d3fac4dd3837e93abb779104568b94019c9a161752c7ca670c77	eca109f5a12abf929dcf5496235d602252357780e6144331b05c562afe71e3be	628241c3a2688da4957c97a3dac3358948195e08982a7c644332c6f9b4d5fd27
val.jsonl	833829ce30d929a8ce502036e3f7b049ef8279770558a632efaaf6bd9e54cb03	890c0c4a456e1c9d4ed03e12c7d8fbbb95364c00640fb83bf2ee262be022ea01	cf2b87c608293faae065c5256f056fe8cbe924f8161e5ee5c8beb56c0abc9d6b
test.jsonl	66c6318d57d2021a85384020f8fb52f5903f37353ec7593e3e6b07c6215dbac1	1977742fe562415fbfb6dc76f2dceb9a81ba0d737e6bf6caec18beed698a2dcd	203dc715d79247c46476b63288cb0b453c7ec33cb9c94691699533509c76fac6

Each column is one bundle directory; the 0.82 column is the repository root. A per-version SHA256SUMS lists only the four files in that column.

6. Intent Recall and Dynamic Training-Card Construction

This section describes how the dataset was constructed. It is the dataset-construction recall pipeline, not a claim about every production request path. The construction reuses the production card renderer and the production recall SQL shape, and deliberately disables production capability-availability filters so that all 18 capabilities participate.

6.1 Construction pipeline

Stage 1 — data preparation, split and recall

Stage 2 — candidate filtering, dynamic option construction and export

Stage 2 continues from the recall scores produced by Stage 1, which the per-capability maximum score consumes.

Mermaid source

Stage 1

flowchart TD
    A["Source CSV<br/>(human-annotated data)<br/>49,241 rows"] --> B["Eligibility and deduplication<br/>48,333 rows"]
    B --> C["Family-aware split<br/>train 36,197 / val 2,407 / test 9,729"]
    C --> D["Query embedding<br/>BGE-M3 INT8, 1024-d, L2-normalised"]
    D --> E["Two-lane ANN recall per query<br/>standard lane: 64 / similar lane: 64"]

Stage 2

flowchart TD
    F["Per-capability maximum score<br/>score = 1 - cosine distance"] --> G{"max score >= threshold?"}
    G -->|yes| H["Deterministic sort<br/>(-score, capability_id)"]
    G -->|no| I["Not a candidate"]
    H --> J["Top-3 at most, never padded<br/>K = number of selected capabilities"]
    J --> K["K capability cards + 1 REJECT card<br/>options = K + 1, 1..4"]
    K --> L["In-place K=0 to K=1 repair<br/>action rows with a valid gold"]
    L --> M["Deterministic option shuffle<br/>reject-shuffle-v1 over the row id"]
    M --> N["Four-field row projection<br/>message_id / language / input / label"]
    N --> O["Frozen JSONL row<br/>train / val / test"]

Construction order matters: the split is assigned before embeddings and recall, and the same split is reused by all thresholds. Thresholds are applied after recall, to the per-capability maximum scores, so the three versions do not re-run retrieval.

6.2 Annotated query

The source CSV carries the request and its gold annotation. A real source row used in the default bundle:

Source field	Value
query (becomes input.query)	Can you say “Labels must be supported by the original sentence. Please confirm whether the final output preserves every citation.” in Chinese?
language	en (resolved)
is_action	true
capability_id	cap-translate

The source annotation determines the gold label, but it is not shown to the model. The model sees only the request plus the constructed candidate cards. The source row's candidate list does not exist: candidates are created by recall at construction time.

Language resolution is explicit-first: an explicit zh or en source value is kept; otherwise the query script is inspected. Han only → zh, Latin only → en, Han plus Latin → mixed; mixed and unknown are never folded into zh or en.

Eligibility removes empty queries, invalid is_action, action rows without a valid capability, no-action rows carrying a capability, normalized duplicate queries, and rows whose normalized text is itself a standard or similar retrieval-corpus query. The last exclusion prevents the evaluation sample from being its own retrieval answer.

6.3 Vector recall

Queries are embedded with the frozen embedding profile eric-ml-nlp/Delveta-BGE-M3-v1#INT8 through the TEI service: 1024 dimensions, L2-normalized vectors. Recall uses the same two-lane shape as the production retriever:

Lane	Corpus table	Pool per query	Score
standard	capability_standard_queries	64	1 - cosine_distance
similar	capability_similar_queries joined to its standard query	64	1 - cosine_distance

Each lane returns its nearest 64 corpus rows. The score is computed by pgvector's cosine distance operator (<=>), converted to similarity by 1 - distance; scores are stored unrounded, so threshold comparisons are not affected by display rounding. Both query and corpus vectors are L2-normalized, so this is cosine similarity.

The ANN index is approximate (HNSW), not an exhaustive exact-cosine scan. The build record contains an exact-cosine coverage audit for this reason; the dataset does not claim that approximate retrieval is identical to exact retrieval.

The build stores retrieval hits at or above an evidence floor of 0.80. Every published threshold is at least 0.82, so a discarded hit could not have passed any threshold; the floor cannot change the candidate sets in this release.

Business filters (enabled, status, intent_kind, UI visibility) are intentionally not applied in this experiment: all 18 capabilities, including cap-research, participate. Production routing may apply additional filters that this dataset does not model.

6.4 Thresholding and capability aggregation

For each query, all hits from both lanes are reduced to one maximum score per capability. Duplicate hits for the same capability — from the standard lane, the similar lane, or multiple corpus rows — never create duplicate candidate cards.

A capability becomes a candidate when:

max_score(capability) >= threshold

The comparison is inclusive. Thresholds are 0.82 (default), 0.85 and 0.88. Candidates are sorted deterministically by:

(-score, capability_id)

The first key is descending score; equal scores are broken by ascending capability id. The sort does not use corpus id, lane or insertion order, so the candidate order is reproducible.

6.5 Top-3, K, and the final option count

After thresholding, at most three candidates are kept:

retrieved = [capabilities with max_score >= threshold]
top3      = retrieved[:3]
K         = len(top3)
options   = K capability cards + 1 REJECT card

• K counts capability cards only. REJECT is not part of K.
• No padding. If only one capability passes the threshold, the row has two options (one capability plus REJECT), not four. If none pass, the row has one option (REJECT).
• Top-3 is a cap, not a target. Many rows have K=1 or K=2; the higher the threshold, the more rows fall below three candidates.
• OPTION_SLOTS = 4 is only the maximum. It is the adapter's upper bound, not a requirement that every row have four options.
• The recalled Top-3 is not the final card count. The final card list is the recalled Top-3 (zero to three capabilities) plus the REJECT card. The option list is then shuffled, so its stored order is not the recall ranking.

The training sentence is built from input.options in their stored post-shuffle order; label points at the gold slot within the actual K + 1 options.

6.6 REJECT construction and labels

Every row has exactly one REJECT option with this frozen card text:

### REJECT
tool: none
does: No listed capability correctly handles the user's request. Choose this option only when none of the other options is the right capability.

中文：以上列出的能力都不适用于该请求。只有当其它选项都不是正确的能力时才选它。

Label precedence is evaluated in this order:

Condition	Construction label	Provenance (audit only)	Training row
Source row is No-Action	REJECT	no_action	label points at the REJECT option
K = 0	REJECT	no_candidates	label points at the REJECT option
Source gold capability is in Top-3	that capability	gold_in_candidates	label points at that capability option
Gold is valid but not in Top-3	REJECT	gold_not_retrieved	label points at the REJECT option

No-Action is checked before K=0, so a No-Action row with no candidates is recorded as no_action, not no_candidates. The three reject provenances are kept separate in the construction audit; they cannot be recovered from a training row, which stores only the label index of the REJECT option.

In-place K=0 → K=1 repair. Source action rows whose gold is valid but which had no candidate above the threshold would otherwise have a single REJECT option with no capability to learn from. The official repair injects the source gold capability as the only capability candidate:

• keep REJECT; K=1; options=2;
• base order is [gold capability, REJECT], then shuffled with seed prefix k0-to-k1-inplace-v1:<threshold>:<split> over the row id;
• no retrieval score exists for the injected candidate; the audit marks it candidate_origin = "source_gold_injected" and retrieval_score_available = false;
• the pre-repair retrieval evidence remains in the construction audit and is not rewritten.

Post-repair, every remaining K=0 row is a source No-Action row. Repaired counts are 152 (0.82), 603 (0.85) and 2,466 (0.88) across the full 48,333 rows.

6.7 Dynamic card example

A real row from the default training split (threshold 0.82):

{
  "message_id": "3fe59439-e944-4bef-ad64-716e0d922944",
  "language": "en",
  "input": {
    "query": "Can you say “Labels must be supported by the original sentence. Please confirm whether the final output preserves every citation.” in Chinese?",
    "options": [
      "cap-translate: ### cap-translate\ntool: translate\ndoes: Translate provided text from one language into another. Use it when the user explicitly asks to translate text into a different language. It does not apply to summarizing, explaining, or rewriting text in the same language.\n\n中文：把给定文本从一种语言翻译成另一种语言。当用户要求“翻译成某种语言”时适用，附件或引用中的文字同样可以翻译。要求总结、解释或同语言改写的，都不属于它。\nnegative examples:\n- (none curated)\nparams:\n- text (string, required): the text to translate, copied from the sentence\n- target_language (string, optional): optional target language name (e.g. \"English\", \"Chinese\"); omitted -> executor default (English)",
      "cap-pdf-extract-text: ### cap-pdf-extract-text\ntool: pdf_extract_text\ndoes: Extract the complete raw text from an attached or resolved document. Use it when the user asks to extract or export the document's full text. It does not apply to extracting only tables or to reading and understanding the document for question answering.\n\n中文：提取附件/已解析文档的完整原始文字内容。当用户要求提取全文或导出文档原始文字时适用。只转换表格的应交给表格转换；要求理解、阅读并据此回答问题的，不属于它。\nnegative examples:\n- (none curated)\nparams:\n- pages (string, optional): optional page spec like \"2\" or \"1-3\" from the viewer context; omitted = whole document\n- asset_id (string, required, max_len=64): asset id of the attached/opened document (from turn facts, not the sentence)",
      "REJECT: ### REJECT\ntool: none\ndoes: No listed capability correctly handles the user's request. Choose this option only when none of the other options is the right capability.\n\n中文：以上列出的能力都不适用于该请求。只有当其它选项都不是正确的能力时才选它。"
    ]
  },
  "label": 0
}

K = 2, len(input.options) = 3, and label = 0 selects cap-translate. The first two options are the two capabilities that passed the threshold; the third is the mandatory REJECT. The cards are rendered in the B_noprov view. The first option text is:

### cap-translate
tool: translate
does: Translate provided text from one language into another. Use it when the user explicitly asks to translate text into a different language. It does not apply to summarizing, explaining, or rewriting text in the same language.

中文：把给定文本从一种语言翻译成另一种语言。当用户要求“翻译成某种语言”时适用，附件或引用中的文字同样可以翻译。要求总结、解释或同语言改写的，都不属于它。
negative examples:
- (none curated)
params:
- text (string, required): the text to translate, copied from the sentence
- target_language (string, optional): optional target language name (e.g. "English", "Chinese"); omitted -> executor default (English)

The stored option order is the post-shuffle order; it is not evidence that cap-translate scored highest. Per-candidate scores and retrieval rank exist only in the construction audit, not in the training row.

6.8 Leakage protection

Risk	Protection
The evaluation query is also a retrieval-corpus query	Exclude source rows whose normalized text equals a standard or similar corpus query (908 excluded: 36 standard, 872 similar)
Paraphrased members of one family cross splits	Assign whole query families to one split; families are indivisible units
The card view names the answer	Store the B_noprov view: query examples: and the evidence: provenance line are removed
The model learns a slot prior	Shuffle all K + 1 options per row with a deterministic row-keyed seed
Retrieval rank leaks through option order	Store only the shuffled option list; scores and ranks are not training fields
Test data influences training or model selection	The training and selection stages read validation only; test is benchmark-only
Calibration rows receive gradient updates	Exclude the frozen 400 calibration rows from the effective train set

The published JSONL files were checked for cross-split overlap by message_id, exact input.query, normalized input.query, rendered options plus label, and the construction query-family heuristic: no overlap was found. There are 48,333 unique normalized queries and 36,621 unique family signatures across the release.

Limitation: the source has no parent/derived-record id, so corpus overlap can only be detected by exact or normalized text equality. The dataset does not prove that no semantic paraphrase of a corpus query exists; that remains a documented source-quality limitation.

6.9 Version differences

All three versions derive from one shared recall pass. For a fixed query, the per-capability maximum scores are identical across versions; only the threshold changes:

• lower threshold (0.82) → more capabilities pass → larger K → more capability labels;
• higher threshold (0.88) → fewer capabilities pass → smaller K → more REJECT labels;
• the source population, split assignment, calibration indices and label rules are unchanged.

Consequences for the release:

Version	Mean K	Candidate-labelled rows	REJECT rows	K=3 rows
0.82	1.880206	44,201	4,132	14,322
0.85	1.315871	44,009	4,324	3,270
0.88	1.037200	44,061	4,272	213

The gold label can change even though the source annotation does not: a gold capability that passes at 0.82 may fall below 0.88 and turn the row into gold_not_retrieved (REJECT). Rows that had no candidate are repaired independently per threshold, which is why the repaired counts rise as the threshold rises.

7. Intended Use

Intended use

• fine-tuning and evaluating a Laya-style non-autoregressive capability-selection model;
• studying dynamic option counts, abstention with an explicit REJECT option, and the effect of retrieval thresholds on candidate sets and labels;
• reproducible benchmark comparisons across the three frozen thresholds.

Out of scope

• Not a generalization benchmark: it applies only to this system's vertical capability-selection task.

8. Related Resources

Model, scripts and checkpoints

• Model: eric-ml-nlp/Delveta-LayaChoice-v2
• Training scripts: fine_tuning/laya_finetune/V2
• Checkpoints: eric-ml-nlp/Delveta-LayaChoice-v2-checkpoints

Repository files

• fine_tuning/laya_finetune/V2/README.md — the fine-tuning recipe, token budget, stages and historical run record.
• layachoice_v2_verify.py — the stdlib-only bundle verifier (hashes, schema, labels, splits, calibration, threshold consistency).
• layachoice_v2_dataset.py — the adapter and loader that define what a valid row is.
• data_versions/dataset_construction/ — the V4 construction bundle: source scripts, build records, audits and reports.
• capability_ids.json — the 18 capability ids.
• Base checkpoint: convaiinnovations/laya, subfolder multilingual, revision 55cf4c4ebb4ebe31b2550e8bdf3bd21b99753851.
• Embedding model: eric-ml-nlp/Delveta-BGE-M3-v1, INT8 profile, 1024 dimensions.

9. License and Citation

License. The Delveta repository code is licensed under the GNU Affero General Public License v3.0. No separate license has been declared for this dataset content. This card does not grant a dataset license; until the maintainer declares one, do not assume that the data may be redistributed or used commercially. The source data's own terms also apply.

Citation. There is no associated paper, DOI or formal citation for this dataset. If a reference is required, cite the dataset name and the exact version or revision used, for example:

Delveta. Delveta-LayaChoice-v2-Data. Version recall_threshold_082 (or the specific
threshold directory), <revision>.

Do not cite a paper that does not exist. The frozen checksums in SHA256SUMS identify the exact data bytes used in an experiment.