cbuk10001/Cam1-training-data

Cam1 original training data

This is the actual 20,000-record training set used to produce Cam1, including the cached teacher distributions used on a fixed 2,000-record subset. It is a derived collection of public-source annotations and controlled synthetic decision tasks. It is not the Qwen base model's pretraining dataset and is not the later Cam1-v2 curriculum.

The source labels are preserved. Teacher outputs supplement them and are not represented as human ground truth. The collection supports a custom joint-schema decision head with Choice, Noul (boolean), and Score (ordered-level) judgments.

Contents and format

The default Parquet configuration has exactly 20,000 unique training IDs. Its columns are id, source, source_license, group_id, split, has_teacher_probabilities, and record_json. record_json is a lossless JSON serialization of the actual heterogeneous native training record; it includes state, questions, labels, semantic mappings, provenance, and teacher fields where available. Parse this column to recover the original schema rather than interpreting it as chat text.

import json
from datasets import load_dataset

data = load_dataset("cbuk10001/Cam1-training-data", split="train")
record = json.loads(data[0]["record_json"])

For byte-exact reproduction, original/train.records.jsonl.gz contains the original hard-label data and original/train.teacher.records.jsonl.gz contains the actual LoRA training input with cached teacher fields. Decompression yields the original bytes. training-manifest.json pins compressed and uncompressed hashes, source revisions, and counts. The viewer's JSON serialization is semantically identical; it is not claimed to preserve original whitespace.

Source	Records	Mapping	Original license
Stanford SNLI	8,000	Entailment/neutral/contradiction Choice and support Noul	CC BY-SA 4.0
PolyAI BANKING77	5,000	Intent Choice with explicit no-match	CC BY 4.0
Google Research GoEmotions	5,000	Independent Noul emotion judgments	Apache 2.0
LocalLLaMA Typed Decisions	960	Existing Choice/Noul/Score workflows	Apache 2.0
Project controlled synthetic	1,040	Rule-labeled functional availability, support, and evidence	Apache 2.0

Preparation and actual use

Preparation used seed 3407. Source/document groups were partitioned before schema augmentation. Exact and lexical near-duplicate grouping used normalized word-set Jaccard similarity at least 0.90, transitive union, and source-group closure. Official test groups took precedence. One overlapping Typed Decisions training item was excluded before allocating the 960 records. These checks do not establish freedom from semantic or base-model pretraining overlap.

Training augmented question and option IDs/order, semantic wording, number of judgments, and candidate subsets while preserving label mappings. The model receives decision evidence and question definitions; provenance and gold labels are supervision/metadata rather than evidence. Training used primitive weights. The validation selection objective balanced primitives and sources within each primitive.

The original workflow kept 565 validation, 1,006 calibration, and 11,028 evaluated test records separate. This public dataset includes training only: none of those held-out examples, official evaluation corpora, or Cam1-v2 sealed tests are included. Complete private audit/recovery records preserve the original split manifests.

The teacher was the original local Clef implementation/model at revision 17f0b0ad64efb65d273590632833508766b2aae6, evaluated in BF16 on a deterministic 2,000-record training subset selected with seed 3407. Teacher distributions coexist with source labels. Cam1's selected LoRA stage used the 20,000-record teacher-augmented file with distillation weight 0.25; no additional teacher calls occurred for this publication.

Attribution, licenses, and limitations

This is a mixed-license dataset. Each source retains its own license and attribution; the source_license column and LICENSE identify the applicable terms. The project does not relicense the upstream records under a blanket model license. The derived SNLI portion retains CC BY-SA 4.0. Project-created controlled synthetic examples and original schema/preparation contributions are provided under Apache 2.0, subject to the underlying source terms where applicable.

Pinned sources:

• Stanford NLP, SNLI, revision cdb5c3d5eed6ead6e5a341c8e56e669bb666725b.
• PolyAI, task-specific-datasets/BANKING77, revision 57ec275d8078af65b7731c2a98be812d844a6d6b.
• Google Research, GoEmotions, revision add492243ff905527e67aeb8b80c082af02207c3.
• LocalLLaMA, Typed Decisions, revision d0e2f0c42fef86cc15d1688d25a19f5ba7c85b18.

Source text and annotations may be noisy, biased, or offensive. Synthetic labels are not human-reviewed truth. The small source mixture and templated cases do not establish performance across domains or languages. No private customer data was introduced; records derive from the identified public datasets and controlled generation. Model quality and inference limitations are described in the Cam1 model card, and official public evaluation outputs are at Cam1-decision-index.