README.md

---
license: other
task_categories:
  - text-generation
language:
  - en
tags:
  - grammatical-error-correction
  - language-learning
  - distillation
size_categories:
  - 1K<n<10K
---

# qninhdt/lexi-grader-sft

Teacher-generated training data for a **sentence grader**: given a learner's
English sentence and one specific dictionary sense of a target word, produce an
inline correction, a meaning band (0-4), and one line of feedback.

This card is generated from the pipeline's own report JSON. Every number below
was measured by the run that produced the data.

## What is in here

- **1636** generated learner sentences (`raw/raw_texts.parquet`)
- **1578** accepted gradings (`raw/raw_labels.parquet`)
- **190** distinct dictionary senses

### Splits

| Split | Rows |
|---|---:|
| `train` | 1225 |
| `val` | 179 |
| `test` | 174 |

Grouped by target word, not by row. Sentences appearing in more than one split: **0**. Rejected during validation: **0** of 1578.

```json
{
  "correction": "He [speak>speaks:agr] very [eloquent>eloquently:form].",
  "meaning": 3,
  "feedback": "Right sense, but the verb needs to agree with the subject."
}
```

| Operation | Syntax | Example |
|---|---|---|
| Replace | `[A>B:tag]` | `[speak>speaks:agr]` |
| Delete | `[A>:tag]` | `the [the>:art] very` |
| Insert | `[>B:tag]` | `went [>to the:art] store` |

A clean sentence is re-emitted verbatim. An unreadable one yields
`correction: null`. `grammar` and `naturalness` are **computed from the
correction's tags** by the formula in `band_config.json`, not generated by the
model — so identical error sets always score identically, and thresholds stay
retunable without regenerating anything.

## How it was built

Two calls, deliberately separated:

1. **Diversifier** — knows a spec (learner profile, target band, error recipe)
   and writes learner-like text. The spec is a diversity knob and **never a
   label**.
2. **Grader** — sees only `{target, sense, text}` and produces the answer. This
   is byte-for-byte the prompt the student model runs at inference.

Single-call self-labelling was rejected: it produces labels that describe the
instruction rather than the text, and the defect is invisible afterwards because
a correct row and a wrong one look identical.

| | |
|---|---|
| Teacher model | `claude-opus-5` |
| Endpoint | `https://api.vilao.ai/v1` |
| Call 1 requests | 190 |
| Call 2 requests | 190 |
| Format validity | 0.6947 |
| Batch diversity (distinct-2) | 0.9555 |

The teacher model string is the one the endpoint reported. It was reached
through an OpenAI-compatible proxy, so it identifies the endpoint's advertised
model rather than an independently verified checkpoint.

## Distribution

| `meaning` | Rows | Share |
|---|---:|---:|
| 0 | 143 | 9.1% |
| 1 | 353 | 22.4% |
| 2 | 114 | 7.2% |
| 3 | 271 | 17.2% |
| 4 | 697 | 44.2% |

Middle bands {1,2,3} hold **0.4394** of rows.

### Error tags

| Tag | Count |
|---|---:|
| `art` | 617 |
| `agr` | 428 |
| `tense` | 402 |
| `num` | 357 |
| `prep` | 351 |
| `form` | 274 |
| `word` | 174 |
| `pron` | 112 |
| `other` | 96 |
| `order` | 90 |
| `punc` | 85 |
| `coll` | 84 |
| `poss` | 48 |
| `unnat` | 47 |
| `sp` | 39 |
| `part` | 25 |

## Quality gates

| Gate | Value | Threshold | Blocking | Result |
|---|---:|---:|---|---|
| `G1_self_consistency` | 0.9679 | 0.7 | yes | pass |
| `G2_band_coverage` | 0.4677 | 0.4 | yes | pass |
| `G3_format_validity` | 0.9645 | 0.9 | no | pass |
| `G4_batch_diversity` | 0.9555 | 0.7 | no | pass |
| `G5_other_tag_share` | 0.0297 | 0.05 | no | pass |
| `G6_batch_single_parity` | 1.0 | 0.8 | no | pass |

Overall: **pass**; blocking gates: **pass**.

### Teacher self-consistency (the ceiling on any student)

| Measure | Value |
|---|---:|
| `meaning` QWK | 0.967852 |
| `correction` edit-F1 | 0.857129 |
| Rows re-graded | 197 |

The same sentences were graded twice, blind, with the cache disabled. A
student cannot exceed the agreement its teacher has with itself, so report
fidelity against these numbers rather than against 1.0.

## Limitations

- **No human gold set.** Every label is a teacher's opinion. Nothing here is
  verified against ground truth.
- **The real-learner distribution is unverified.** Sentences are model-written
  imitations of learner errors, not collected from learners.
- **Bands are uncalibrated** until `lexi data calibrate` has run;
  `band_config.json` carries `"calibrated": false` when that is still true.
- **`feedback` is unmeasured.** No metric in this snapshot evaluates it.
- **The error mix is skewed relative to human-annotated learner text.** Against W&I+LOCNESS train split (18,224 rows, the same 16 tags), this data over-produces `agr` (13.2% vs 2.8%), `num` (11.1% vs 4.8%), `poss` (1.5% vs 0.7%) and under-produces `punc` (2.6% vs 15.4%), `sp` (1.2% vs 8.5%), `unnat` (1.5% vs 9.0%), `word` (5.4% vs 17.1%). The cause is the learner profiles: their `error_bias` fields drive what call 1 writes, and `punc`, `sp`, `pron`, `poss` and `other` appear in no profile at all. Punctuation and spelling are among the commonest real learner errors, so a model trained on this alone will be weakest there. Full comparison in `reports/tag-distribution-reference.json`.
- **`grammar` and `naturalness` are uncalibrated.** `band_config.json` carries `"calibrated": false`, so those two bands come from the shipped design guesses rather than from this corpus's penalty distribution. `meaning` is unaffected — it is the teacher's own answer. Calibration on this corpus produced duplicate cut points, because 81% of rows carry no `usage` error at all and no threshold exists inside that mass of zeros; a five-band usage scale is not supported by data this clean.

## Not included

The stage-A correction-format data converted from W&I+LOCNESS is **not** part of
this dataset. That corpus's licence forbids redistributing any part of it to a
third party, so it stays local to the machine that built it.

## Citation

```bibtex
@misc{lexi_grader_dataset,
  title  = {qninhdt/lexi-grader-sft},
  note   = {Teacher-generated sentence-grading dataset},
  year   = {2026},
}
```