README.md

---
license: cc-by-4.0
task_categories:
- text-generation
tags:
- structured-generation
- information-extraction
- lora
pretty_name: OO-LD quantity extraction, lean prompt
---

# oold-quantities-lean

What every adapter under this organisation was trained on. Chat-formatted
examples of quantity extraction, generated from the
[quantity schemas](https://huggingface.co/datasets/OO-LD/oold-quantity-schemas) and rendered to prose.

## The prompt is the evaluation prompt

The system message is built by the same function that builds it at evaluation
time, under the same arm, and is character-identical to it. Trained on one
wording and evaluated on another, a result is a wording difference.

Lean means the class catalogue is shown and the schema is not. Trained with a
full schema in context, a model learns to copy from the prompt, and the
tuned-with-no-schema condition the adapters exist for would measure nothing.

## Held out by class, and by register

Roughly half the describable classes are excluded from training entirely. At
evaluation the offered catalogue still carries both halves, or a held-out
class could not be chosen and the contrast could not be measured.

The register is held out too. These documents are generated prose; the
adapters are scored on Wikipedia sentences. A number measured this way is not
a report on the generator.

Seeds are disjoint from every evaluation draw, so no scored document was
trained on.

## Every example round-trips

Each assistant turn is fed back through the real extractor and the real
grader and must score 1.00. A training set that does not round-trip teaches a
model answers the benchmark would mark wrong. 17 of 1,017 draws were rejected
by that gate.

## Files

| file | examples |
|---|---|
| `train.jsonl` | 1,000 |
| `validation.jsonl` | 200 |

Both carry a UTF-8 byte order mark. Kept rather than stripped, because these
are the bytes the adapters were trained on; read them with `utf-8-sig`.

Regenerate with [oold-llm-bench](https://github.com/OO-LD/oold-llm-bench).