README.md

---
license: cc-by-4.0
task_categories:
- text-classification
language:
- en
tags:
- reinforcement-learning
- calibration
- bandit-feedback
- decision-making
- rlcd
size_categories:
- 10K<n<100K
configs:
- config_name: default
  data_files:
  - split: train
    path: train.jsonl
  - split: validation
    path: val.jsonl
  - split: test
    path: test.jsonl
---

# RLCD Decision Dataset (v1)

Typed decision questions for training and evaluating models that answer with a calibrated probability distribution over declared options instead of generated text. Built for [RLCD — reinforcement learning for calibrated decisions](https://github.com/anthony-maio/eve-rlcd) (see also the trained export [anthonym21/qwen3-0.6b-rlcd-decision](https://huggingface.co/anthonym21/qwen3-0.6b-rlcd-decision)).

Every row is one typed question over a context: a `choice` question over unordered options, a `score` question over ordered levels, or a `noul` (yes/no) question. The RLCD training loop treats each row as a bandit arm — the environment reveals only whether the sampled option was correct, never the label — and the reward `r = c - p_a` pushes the sampled probability toward calibration.

## Schema

| field | type | meaning |
|---|---|---|
| `primitive` | string | `choice` (unordered), `score` (ordered levels), `noul` (yes/no) |
| `context` | string | the state / passage the question is about |
| `question` | string | the question, in words |
| `choices` | list[string] | the declared options (2–26) |
| `ordered` | bool | whether the options are ordered levels (score questions) |
| `answer` | int | index into `choices` of the correct option |
| `source` | string | which of the 8 sources the row came from |
| `id` | string | stable row id `<source>-<n>[-<aspect>]` |

One JSON object per line; CRLF line endings preserved from the release build.

## Splits

| split | rows | sources | per-source |
|---|---|---|---|
| train | 64,000 | 8 | 8,000 |
| validation | 8,000 | 8 | 1,000 |
| test | 8,000 | 8 | 1,000 |

Train primitives: 34,667 choice / 18,667 score / 10,666 noul. Validation and test: 4,334 / 2,333 / 1,333 each. `stats.json` (included) has the full per-source breakdown.

## Provenance

Built with `rlcd.data build --per-source 8000 --seed 0` on 2026-09-17 from the code at commit `57a179b`. These are the exact files used for every published run in the repo's results tables; a later rebuild produced different bytes, so use these for comparable runs.

Source datasets (question converters in [`rlcd/data.py`](https://github.com/anthony-maio/eve-rlcd/blob/main/rlcd/data.py)):

| source | upstream dataset | upstream license | transformation |
|---|---|---|---|
| bitext | [bitext/Bitext-customer-support-llm-chatbot-training-dataset](https://huggingface.co/datasets/bitext/Bitext-customer-support-llm-chatbot-training-dataset) | CDLA-Sharing-1.0 | intent classification over sampled label subsets |
| banking77 | [legacy-datasets/banking77](https://huggingface.co/datasets/legacy-datasets/banking77) (mirror of PolyAI/banking77) | CC-BY-4.0 | intent classification over sampled label subsets |
| ag_news | [fancyzhx/ag_news](https://huggingface.co/datasets/fancyzhx/ag_news) | unspecified upstream | topic classification, fixed 4 options |
| mnli | [nyu-mll/multi_nli](https://huggingface.co/datasets/nyu-mll/multi_nli) | CC-BY-3.0 / CC-BY-SA-3.0 (mixed) | premise–hypothesis relation, fixed 3 options |
| sst5 | [SetFit/sst5](https://huggingface.co/datasets/SetFit/sst5) | unspecified upstream (SST derivatives) | 5-level sentiment, ordered |
| yelp | [Yelp/yelp_review_full](https://huggingface.co/datasets/Yelp/yelp_review_full) | Yelp Dataset terms | 5-level star rating, ordered, text truncated to 1,500 chars |
| boolq | [google/boolq](https://huggingface.co/datasets/google/boolq) | CC-BY-SA-3.0 | yes/no question over passage |
| triage | synthetic generator in `rlcd/data.py` (original) | — | enterprise support tickets with department / priority / escalation questions |

The `triage` rows are fully synthetic (original work). All other rows are transformed subsets of the upstream datasets above; credit for the underlying texts belongs to the upstream sources, and their terms (some share-alike) apply to those portions. This repo is distributed as CC-BY-4.0 as a convenience tag; if your use is sensitive to the upstream terms, follow the links and check them.

## Checksums (md5)

```
train.jsonl  cdfee4c9792751cf5b22668eb3f9dc33
val.jsonl    4b95911ffc76ed1789f7989f623d0a5a
test.jsonl   15c33b165d70639d8bd7d23d624908f4
```

## Intended use

Research on decision-making LLMs under bandit/outcome-only feedback, probability calibration (ECE, Brier), and confidence-aware classification. Not a benchmark of world knowledge: every split is in-distribution for the sources above and the questions are template-generated.

## Citation

```bibtex
@software{maio2026eve_rlcd,
  title = {eve-rlcd: reinforcement learning for calibrated decisions},
  author = {Anthony Maio},
  url = {https://github.com/anthony-maio/eve-rlcd},
  year = {2026}
}
```