Instructions to use mydeskmate/grammar-v7 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use mydeskmate/grammar-v7 with Transformers:
# pip install -U transformers accelerate # Load model directly from transformers import AutoTokenizer, AutoModelForSeq2SeqLM tokenizer = AutoTokenizer.from_pretrained("mydeskmate/grammar-v7") model = AutoModelForSeq2SeqLM.from_pretrained("mydeskmate/grammar-v7", device_map="auto") - Notebooks
- Google Colab
- Kaggle
grammar-v7 (run #1 exam-push) -- FAILED the do-no-harm gate; NOT shipped
Educational / provenance release. This checkpoint is not used in
production. The live site keeps shipping mydeskmate/grammar-v5.
What it was
Run #1 of the GPU exam-push recipe: warm start from grammar-v5, 1 epoch on
the standard mix (coedit + jfleg x4 + agentlans x0.8), 1250 steps on a
Google Colab Tesla T4, official scorers, pre-registered selection rule.
Measured (official scorers, GPU run)
| benchmark | v5 (live) | this run | target | verdict |
|---|---|---|---|---|
| CoNLL-2014 test F0.5 | 47.35 | 52.10 | >= 65 | improved (+4.75) |
| BEA-2019 dev F0.5 | 33.90 | 35.48 | >= 70 | improved (+1.58) |
| JFLEG test GLEU | 62.78 | 63.25 | >= 62.30 | pass |
| Do-No-Harm (clean touched) | 2.39% | 5.00% (25/500) | < 5% | FAIL |
Verdict
The pre-registered rule (CoNLL/BEA up AND GLEU >= 62.30 AND Do-No-Harm < 5%)
failed on the last condition by one sentence of 500 (24/500 would have
passed). Reported as measured; nothing hidden. The checkpoint is kept as the
warm start for run #2, the do-no-harm calibration -> mydeskmate/grammar-v7b.
- Downloads last month
- 32
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support