insight-research/oci-topic-labels-en

OCI Topic Labels (EN)

A seed dataset: 1,779,254 English texts from 47 sources (44 public Hugging Face datasets), each with one topic label (112 labels) assigned by the Oracle OCI Language pretrained text classification service. It was built by taking advantage of available Oracle Cloud credit, and it is a topic-labelled source sample, nothing more. It is a seed to develop and tune methods, not training data: we use it to refine how we classify, extract key phrases and entities, and order items, so that those methods can later be applied to the data we do train on. It is not intended for training. The labels are machine output, not human annotation, and they were not corrected.

The texts are not ours. Each one keeps the license of the dataset it came from; see Sources and Licensing.

Motivation

It all started with one idea: instead of training a model at its final size, grow it in stages. Staged Training for Transformer Language Models (Shen et al., 2022) trains a small model, starts a new stage "when training efficiency starts decreasing", and grows it by doubling its width or its number of layers. Curriculum-Guided Layer Scaling (Singh et al., ICML 2026) goes in the same direction: it adds layers, trains the new ones first with the old ones frozen, then unfreezes the whole model and moves each stage to harder data. That is the plan behind this work: a model that doubles at each step.

The second reason comes from where we start. Our first models are very small, around 2.5M parameters, so the data has to be designed for those sizes and for each one that comes after. A tiny model can't take everything at once; it needs the data grouped and ordered. That is what sent us to look at how others organize training data: by topic and format, as in WebOrganizer (Wettig et al., ICML 2025); mixing by topic works better than mixing by source (Peng et al., 2025); the best mixture or filter changes with model scale (AutoScale, Kang et al., COLM 2025; BETR, Mizrahi et al., 2025; Shukor et al., 2025); and ordering the data adds to selecting it (Beyond Random Sampling, EACL 2026).

So this dataset is not training data, and it is not aimed at training even though the idea was born from it. It is a seed: a place to tune the methods (topic classification, key phrases, entities, and how to group and order items) that we will later apply to the data we do train on. These labels are the first step: one topic for every item of 47 sources, to try out how to group them by theme, order them and build the mix for each stage of the ladder. We also assumed OCI Language would give us key phrases and entities on every document, and it does produce them. But we chose to go straight to classification, which is what the ordering needs, and only requested phrases and entities in an early pilot of about 8,100 documents. Looking back, that may have left the result unbalanced, since the extra signals cover only 0.5 % of the data. Those credits were already spent, so the pilot's phrases and entities are published as a separate small file, not as columns.

Dataset structure

Files:

• data/<source>.parquet: 47 files, one per source, about 440 MB in total.
• pilot/phrases_entities.parquet: key phrases and entities for 8,111 documents of the early pilot, kept apart from the main files (see Pilot).
• LICENSE.md: provenance and license notice, with the declared license of each source.
• sources.json: source id, original dataset, record type, document count, declared license and flags.
• source_specs/<source>.yaml: the spec of each of the 47 sources, with its URL and column mapping (see Source specs).
• taxonomy.json: the 112 labels, each with a definition.
• metadata.json: column descriptions, service, processing dates.
• stats.json: the numbers behind the charts.
• images/: the charts shown below.

Columns:

Column	Description
id	Document id in the original dataset: <source>/<path in the original HF repo>#<row index>. Drop the source prefix to find the file in the original repo.
source	Source id; sources.json maps it to the original dataset.
type	Record type of the source (raw_document, instruction, tool_use, choice_decision, ...).
text	The text that was sent to OCI, cut at 2,000 characters.
text_format	Layout of text: plain, dialogue (lines of role: content) or json (a JSON value, an object or a quoted string; decode it before use, texts cut at 2,000 characters may not parse).
truncated	True when the text was cut at 2,000 characters.
char_count	Characters of the original text, before the cut. It can exceed the length of text by a few characters when surrounding whitespace was trimmed, even if truncated is false (618 documents).
classification	List of {label, score} returned by OCI Language, highest score first. Labels are Parent/Sub paths of the OCI taxonomy.
top_label	The first label of classification.
top_score	Score of top_label, between 0 and 1.

Example row (from banking77):

{
  "id": "banking77/test.jsonl#112",
  "source": "banking77",
  "type": "choice_decision",
  "text": "I'm trying to figure out the current exchange rate.",
  "text_format": "plain",
  "truncated": false,
  "char_count": 51,
  "classification": [{"label": "Finance/Investing", "score": 1.0}],
  "top_label": "Finance/Investing",
  "top_score": 1.0
}

Sources

The license column is what each source declares, collected on 2026-11-08. Check the source yourself before using a text; this is not legal advice. The full evidence is in the project repository (.research/20261108-oci_topic_labels_licensing-research.yaml).

Source	Original dataset	Type	Docs	Declared license
aegis_ai_content_safety	nvidia/Aegis-AI-Content-Safety-Dataset-2.0	boolean_decision	20,188	CC-BY-4.0 (contains offensive content)
agentic_tool_call_12k	pyromind/agentic-tool-call-dataset-12k	agentic_trajectory	4,988	Apache-2.0 (derived from upstream sets)
apigen_function_calling	argilla/apigen-function-calling	tool_use	58,512	CC-BY-4.0
banking77	mteb/banking77	choice_decision	2,282	MIT
belebele	facebook/belebele	choice_decision	488	CC-BY-SA-4.0 (intended as a test set)
berkeley_function_calling_leaderboard	gorilla-llm/Berkeley-Function-Calling-Leaderboard	tool_use	3,101	Apache-2.0
bespoke_stratos_17k	bespokelabs/Bespoke-Stratos-17k	instruction	2,008	Apache-2.0
civil_comments	google/civil_comments	score_decision	168,613	CC0-1.0 (toxic content)
cosmopedia_v2_en	HuggingFaceTB/smollm-corpus	raw_document	31,951	ODC-BY (Common Crawl terms also apply)
culturax_en	uonlp/CulturaX	raw_document	119,417	No license field; "mC4 and OSCAR terms"; gated
dclm_baseline_en	mlfoundations/dclm-baseline-1.0	raw_document	99,341	CC-BY-4.0
finemath_en	HuggingFaceTB/FineMath	raw_document	77,872	ODC-BY (Common Crawl terms also apply)
finephrase_en	HuggingFaceFW/finephrase	raw_document	67,444	ODC-BY (Common Crawl terms also apply)
fineweb_edu_en	HuggingFaceFW/fineweb-edu	raw_document	72,597	ODC-BY (Common Crawl terms also apply)
fineweb_en	HuggingFaceFW/fineweb	raw_document	149,301	ODC-BY (Common Crawl terms also apply)
glaive_function_calling_v2	glaiveai/glaive-function-calling-v2	tool_use	46,295	Apache-2.0
gliclass_v2_0	knowledgator/gliclass-v2.0	choice_decision	14,088	Apache-2.0
gliclass_v2_0_binary	knowledgator/gliclass-v2.0	boolean_decision	17,644	Apache-2.0
gliclass_v2_0_rac	knowledgator/gliclass-v2.0-RAC	choice_decision	9,511	Apache-2.0
gliclass_v3_logic	knowledgator/gliclass-v3-logic-dataset	choice_decision	2,759	Apache-2.0
helpsteer2	nvidia/HelpSteer2	score_decision	5,054	CC-BY-4.0
hermes_agent_reasoning_traces	lambda/hermes-agent-reasoning-traces	agentic_trajectory	6,819	Apache-2.0
hermes_function_calling_v1	NousResearch/hermes-function-calling-v1	agentic_trajectory	6,992	Apache-2.0
jev_bench	Praveenrajus/jev-bench	choice_decision	24,301	Mixed; see the source's manifest
jev_bench_binary	Praveenrajus/jev-bench	boolean_decision	39,223	Mixed; see the source's manifest
jev_bench_score	Praveenrajus/jev-bench	score_decision	25,237	Mixed; see the source's manifest
magicoder_oss_instruct	ise-uiuc/Magicoder-OSS-Instruct-75K	instruction	4,533	MIT (notes the OpenAI usage policy)
multi_nli	nyu-mll/multi_nli	choice_decision	39,342	Mixed (cc-by-3.0, cc-by-sa-3.0, mit, other)
nampdn_tiny_codes	nampdn-ai/tiny-codes	instruction	25,354	MIT (gated)
numinamath_cot	AI-MO/NuminaMath-CoT	instruction	46,898	Apache-2.0
nvidia_open_code_instruct	nvidia/OpenCodeInstruct	instruction	72,722	CC-BY-4.0
open_thoughts_114k	open-thoughts/OpenThoughts-114k	instruction	9,018	Apache-2.0
open_web_math	open-web-math/open-web-math	raw_document	109,542	ODC-BY (Common Crawl terms also apply)
opencoder_sft_stage1	OpenCoder-LLM/opc-sft-stage1	instruction	28,148	MIT
prometheus_feedback_collection	prometheus-eval/Feedback-Collection	score_decision	21,739	CC-BY-4.0
prometheus_preference_collection	prometheus-eval/Preference-Collection	choice_decision	19,067	CC-BY-4.0
s1k	simplescaling/s1K	instruction	996	Apache-2.0
simplestories_en	SimpleStories/SimpleStories	raw_document	18,210	MIT
squad_v2	rajpurkar/squad_v2	grounding_qa	17,375	CC-BY-SA-4.0
tinystories_en	roneneldan/TinyStories	raw_document	17,566	CDLA-Sharing-1.0
toolmind	Nanbeige/ToolMind	agentic_trajectory	21,606	Apache-2.0 (folds in other sets)
trace_commons_agent_traces	trace-commons/agent-traces	agentic_trajectory	28	CC-BY-4.0 (compilation only)
typed_decisions	LocalLLaMA/typed-decisions	mixed_decision	1,600	Apache-2.0
ultrafeedback_binarized_preferences	argilla/ultrafeedback-binarized-preferences	choice_decision	18,140	None declared (upstream openbmb/UltraFeedback is MIT)
vitaminc	tals/vitaminc	choice_decision	181,481	CC-BY-SA-3.0
xlam_function_calling_60k	Salesforce/xlam-function-calling-60k	tool_use	31,231	CC-BY-4.0 (gated, "research purposes")
xnli	facebook/xnli	choice_decision	18,632	CC-BY-NC-4.0 (from the GitHub LICENSE; none on the HF card)

Label origin and taxonomy

The labels come from the OCI Language pretrained text classification, called through the oci-ailanguage SDK (2.142.2) from a Bun/TypeScript client. Oracle publishes neither a model version nor label definitions for this service. The service returns Parent/Sub paths; 112 distinct labels appear in the data.

taxonomy.json lists the 112 labels with a definition each. The definitions were written by us, starting from a Gemini draft; they are not Oracle's. Documents were processed from 2026-10-14 to 2026-11-06. Only English texts are included.

Pilot: key phrases and entities

OCI Language can also return key phrases and entities. In the first pilot (2026-10-14) we requested all three features on about 8,100 documents from 10 raw_document sources: cosmopedia_v2_en, culturax_en, dclm_baseline_en, finemath_en, finephrase_en, fineweb_edu_en, fineweb_en, open_web_math, simplestories_en and tinystories_en. After that we ran topic classification only. Because the signals cover 0.5 % of the data, they are not columns of the main files; they are one separate file, pilot/phrases_entities.parquet (config pilot_phrases_entities).

Column	Description
id	Same id as in data/<source>.parquet; join on it to get the text and the topic label. All 8,111 ids are in the main files.
source	Source id.
key_phrases	List of {text, score}, or null if OCI returned none.
entities	List of {text, type, score}, or null if OCI returned none.

8,111 documents: 8,101 with key phrases, 8,086 with entities, 8,076 with both. 357,361 phrases in total (44.1 per document) and 214,044 entities (26.5 per document). The most frequent entity types are QUANTITY (69,152), PRODUCT (54,644), DATETIME (23,296), PERSON (22,316), LOCATION (22,295) and ORGANIZATION (14,379). The file has no text; it carries the same OCI service version as the labels (2.142.2).

Source specs

source_specs/<source>.yaml is a verbatim copy of the spec our project uses for each of the 47 sources. It records where the data comes from and how its columns were mapped. The text of this dataset was built from the columns these specs point to; text_format tells the layout.

Field	Meaning
id	Source id; the same as source in the data.
dataset, url	The original dataset on Hugging Face and its page.
languages	Languages the spec reads.
dataset_type	Record type (raw_document, instruction, ...).
kind, type_file, subsets	How the files are fetched and read: shard kind, file pattern, subsets.
column_mapping	Original column → canonical field, for example question → prompt.
column_mapping_metadata	Original columns kept as metadata.
scan	Which files and text columns our scan reads (for example text_columns).
adapter	Shared reading code, when several specs come from one original dataset (jev_bench, gliclass_v2_0).
gated	Whether the source needs accepting terms before download.
origin	organic, synthetic or mixed.
targets	The model families (LLM, SLM, DM) our project planned that source for. It describes the source, not this dataset, which is not intended for training.

Comments inside the files carry license and row-count notes. Not every field appears in every spec.

Results

All numbers come from this dataset's own labels (stats.json). They describe the distribution of the labels, not their correctness: no human review was done.

raw_document is 43 % of the documents. The largest sources are vitaminc (181k), civil_comments (169k) and fineweb_en (149k).

Parent categories are fairly flat: Computer and Electronics 14.3 %, Science and Technology 13.3 %, Society and State 10.6 %.

The largest sub-labels are Education (8.8 %), Mathematics and Statistics (8.5 %) and Programming (8.0 %).

Code sources are single-topic: magicoder_oss_instruct is 87.5 % Programming, nvidia_open_code_instruct 82.8 % and nampdn_tiny_codes 73.1 %. Math reasoning sets are 54 to 62 % Mathematics. 28 of the 47 sources have a dominant label under 20 %.

Every document has exactly one label (the classification list has length 1), and all 112 labels appear. top_score is bimodal: 48.6 % of the documents score between 0.98 and 1.0, 34.6 % score below 0.5, and the median is 0.945. By type, tool_use has a median of 0.71 and mixed_decision 0.64.

Other (20,863 documents) is a fallback: all its scores are below 0.02. Treat it as unlabeled. It is concentrated: 12,571 of those documents (60 %) come from opencoder_sft_stage1, where 44.7 % of its texts got no topic, and another 3,498 from finemath_en.

The median original length is 564 characters (p95 3,054).

7.5 % of the documents (133,745) were cut at 2,000 characters. By type, the cut falls mostly on agentic_trajectory (68.5 %), then instruction (13.1 %) and raw_document (10.0 %). truncated and char_count keep the original length.

22 labels have fewer than 1,000 documents; the rarest are Gambling (2), Animal Sports (3) and Aging and Geriatrics (5).

Budget and processing

Metered usage was 303.55 USD on OCI AI Language, taken from the Oracle Usage API (settled): 1,224,006 transactions at 0.25 USD per 1,000. Oracle bills ceil(chars per call / 1,000) per call, so short documents batched together cost less than one transaction each, about 0.17 USD per 1,000 documents. It was paid with available Oracle Cloud credit.

Date	Spend (USD)
2026-10-14	4.56
2026-10-16	4.28
2026-10-17	14.27
2026-10-30	4.05
2026-10-31	43.71
2026-11-04	66.19
2026-11-05	51.49
2026-11-06	114.99

The final credit run used 11 blocks (2026-11-04 to 2026-11-06). A full block is about 25 USD and about 156k documents, and took about 60 minutes (around 2,600 documents per minute, 43 per second); the pilot took 75 minutes. Requests carried up to 100 records or 20,000 characters, with about 92 ms latency per call. There were 0 document errors (lost calls were retried) and HTTP 429 stayed under 0.1 % in the later blocks.

The OCI key phrases and entities features were requested only in the early pilot (about 8,100 documents, 0.5 % of the dataset); every other call was topic classification only. The pilot's phrases and entities are in pilot/.

Limitations

• Not intended for training: this is a seed for developing and tuning methods.
• One label per document. Texts that span topics are forced into one.
• Other is a fallback, not a topic; its scores are all below 0.02.
• Scores are not calibrated probabilities, and top_score is bimodal. Do not read it as accuracy.
• The service has no published model version or definitions, so reruns may not reproduce these labels.
• Texts are cut at 2,000 characters; the label reflects the cut text, not the full original.
• 22 labels have fewer than 1,000 documents.
• Sources overlap. The same text can appear in more than one source: 29,929 texts of xlam_function_calling_60k also appear in apigen_function_calling, 4,125 of multi_nli in gliclass_v2_0_binary, 999 of open_web_math in finemath_en. Some sources repeat texts inside themselves (toolmind 1,005, hermes_function_calling_v1 467, opencoder_sft_stage1 189). Nothing was deduplicated; do it before using the texts.
• A pattern scan found e-mail-like strings in about 0.8 % of the documents and phone-like numbers in about 0.5 %, mostly in web crawls and tool-use traces. We did not remove them. The credential-like strings we looked at (AWS key ids, Slack tokens, private-key headers) were examples or placeholders, but we did not check every one.
• Content warning: aegis_ai_content_safety and civil_comments are made of toxic and offensive text, and other sources, web crawls above all, may contain some too. Web sources may contain personal data.
• English only.

Licensing

The labels were produced with Oracle OCI Language, and Oracle's terms for the output of its services apply to them; check those terms before relying on the labels. Each text keeps the license of the dataset it came from, listed per source in Sources and, formally, in LICENSE.md; those licenses were collected from each source's own metadata and may change. We grant no license of our own over the texts or over the Oracle-generated labels; the material we wrote is CC-BY-4.0 (see Author and citation). Users decide what to use and are responsible for checking the source terms.

Author and citation

Nicolas Alejandro Catalogna (insight-research). Prepared with the help of AI tools (Claude and Gemini) for code, drafting and charts; the author directed the work and is responsible for it. Contact, and removal requests from the owners of a source: nicocatalogna@gmail.com. The material written by the author (taxonomy definitions, statistics, charts, this card and the source specs) is released under CC-BY-4.0; see LICENSE.md.

To cite:

@misc{catalogna2026ocitopiclabels,
  author = {Catalogna, Nicolas Alejandro},
  title = {OCI Topic Labels (EN): a seed dataset of topic-labelled texts},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/insight-research/oci-topic-labels-en}}
}

How to load

from datasets import load_dataset

ds = load_dataset("insight-research/oci-topic-labels-en", "default")
one = load_dataset("insight-research/oci-topic-labels-en", "banking77")

The config name is the source id; default loads all 47 files.

pilot = load_dataset("insight-research/oci-topic-labels-en", "pilot_phrases_entities")

This loads the pilot's key phrases and entities; join it with the main files on id.