---
license: apache-2.0
language:
- en
task_categories:
- text-classification
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train.jsonl
  - split: validation
    path: data/validation.jsonl
---

# Commitment statements — synthetic pilot

Small assistant-authored English statements for an experimental binary
commitment/decision classifier. Released for reproducibility, not as an
independently collected benchmark.

- Training: 118 rows, 59 per label.
- Validation/development: 14 rows, 7 per label.
- The original 16-row test set is deliberately not distributed or evaluated.
- Related contrastive families were kept together during the original split;
  all 48 extension rows were appended only to training.
- Exact text, ID and group overlap was checked. Semantic topic independence
  is not guaranteed.

## Labels

`DECISION_EXPRESSED` means an affirmed chosen course of action, firm intention
or commitment. Choosing not to act and explicitly reported past choices count.
`NO_DECISION_EXPRESSED` covers preferences, requests, questions, possibilities,
predictions, example quotations and completed-action reports without an expressed
choice. This is a broad commitment definition, not a complete logging policy.

## Authorship and review

The first pilot contained 100 assistant-authored statements in contrastive
families. An extension added 48 fictional, newly worded statements inspired by
structural patterns in a user-provided decision log. The source log, names,
paths, entry identifiers and project-specific measurements are not included.
The extension was authored after inspecting development errors.

The user accepted proceeding after data validation. This is not a claim of
independent expert annotation or review of each label. Original
`review_status: human-review-pending` fields are retained for provenance rather
than rewritten as expert-approved ground truth.

Extra fields on extension rows, including `settled_outcome_candidate`, are
proposed review annotations, not separately validated training targets. Some
rows lack these fields; consumers should treat them as optional. Training uses
only `text` and `label`.

## Load

```python
from datasets import load_dataset

data = load_dataset("TwilightTechie/commitment-statements-pilot")
print(data["train"][0])
```

## Limitations and intended use

Use for reproducing the small pipeline and learning experiment. Do not use
validation metrics as estimates of real-world reliability. It is short English
text with limited topic and wording diversity; ambiguous, canceled, sarcastic
and multi-turn examples are not qualified. Keep independent real examples for
future evaluation and collect permission before introducing conversation data.

License: Apache-2.0. The repository preserves the Semantic Router license used
for this experiment's source/data context. No license over the private source
log is asserted. Associated model: [`TwilightTechie/vela-commitment-classifier-307m`](https://huggingface.co/TwilightTechie/vela-commitment-classifier-307m).