Add README.md

README.md

@@ -0,0 +1,102 @@
+---
+license: apache-2.0
+language:
+- en
+library_name: transformers
+pipeline_tag: text-classification
+base_model: distilroberta-base
+datasets:
+- aitrainingdatabroker/company-record-types-and-limits
+- aitrainingdatabroker/company-data-licensing-faq-by-topic
+tags:
+- record-types
+- data-licensing
+- text-classification
+- company-data
+- data-export
+---
+
+# Record Type Page Classifier
+
+Reads a passage about company records and says which of 33 record types it is most likely about: support tickets, team chat, CRM history, CAD files, tool exports such as Slack or Jira, and more.
+
+All training text is our own, from pages published at aitrainingdatabroker.com. Nothing comes from a customer.
+
+Related tool: [company-record-types-and-limits-finder](https://huggingface.co/spaces/aitrainingdatabroker/company-record-types-and-limits-finder) (a Space that runs in your browser).
+
+## Why it exists
+
+Before data is prepared, a team has to say what kind of record each file is. This model suggests the record type for a passage so a person can confirm it.
+
+## Use it
+
+```python
+from transformers import pipeline
+
+classify = pipeline("text-classification", model="aitrainingdatabroker/record-type-page-classifier", top_k=3)
+print(classify("A Jira export with issue keys, comments, status changes and linked pull requests."))
+```
+
+## How it was trained
+
+Fine-tuned from `distilroberta-base` for 25 epochs on one RTX 3070. Training data: 419 passages from [company-record-types-and-limits](https://huggingface.co/datasets/aitrainingdatabroker/company-record-types-and-limits), [company-data-licensing-faq-by-topic](https://huggingface.co/datasets/aitrainingdatabroker/company-data-licensing-faq-by-topic), which are our own published page text. It was trained on item descriptions, introductions and section paragraphs from our 33 record-type pages and tested on the questions and answers of those pages, which it never saw in training. The scripts that built the data and trained the model are in `source/`; the full scores are in `scores.json`.
+
+## Test scores
+
+On 107 held-out rows: accuracy 0.533, macro F1 0.497.
+
+| Label | F1 | Count |
+|---|---|---|
+| `accounting-and-erp-records` | 0.0 | 3 |
+| `cad-and-engineering-files` | 0.8 | 3 |
+| `call-recordings` | 0.571 | 3 |
+| `change-orders-and-rfis` | 0.8 | 3 |
+| `claims-and-underwriting-files` | 0.5 | 3 |
+| `client-deliverables-and-drafts` | 0.4 | 3 |
+| `closed-deal-files` | 0.5 | 3 |
+| `code-and-pull-request-history` | 0.0 | 3 |
+| `confluence-data-export` | 0.889 | 4 |
+| `consented-recordings` | 0.667 | 3 |
+| `crm-history` | 0.5 | 3 |
+| `egocentric-video` | 0.857 | 4 |
+| `github-repository-export` | 0.444 | 3 |
+| `gmail-data-export` | 0.0 | 4 |
+| `google-drive-export` | 0.462 | 3 |
+| `hubspot-data-export` | 0.286 | 3 |
+| `inspection-reports-and-photos` | 0.857 | 3 |
+| `jira-data-export` | 0.333 | 3 |
+| `lab-notebooks-and-experiment-records` | 0.667 | 3 |
+| `labeled-tables-and-datasets` | 1.0 | 4 |
+| `microsoft-365-email-export` | 0.364 | 3 |
+| `native-word-excel-and-powerpoint-files` | 0.4 | 4 |
+| `project-and-task-history` | 0.5 | 3 |
+| `quickbooks-data-export` | 0.4 | 3 |
+| `quotes-and-estimates` | 0.667 | 3 |
+| `salesforce-data-export` | 0.0 | 4 |
+| `slack-data-export` | 0.727 | 4 |
+| `sops-and-knowledge-bases` | 0.667 | 3 |
+| `support-tickets` | 0.571 | 3 |
+| `team-chat` | 0.0 | 3 |
+| `theses-and-dissertations` | 0.667 | 4 |
+| `work-orders-and-dispatch-notes` | 0.5 | 3 |
+| `zendesk-data-export` | 0.4 | 3 |
+
+## Limits
+
+- 33 classes with about 13 training passages each, and several are close neighbours (the export guides especially), so treat the top three as candidates.
+- Trained on one broker's page text.
+- This model is not legal advice and not a certification. A person should review its output before anything is shared or decided, and your counsel decides what a rule requires.
+
+## Read more
+
+- [Record types and export guides](https://aitrainingdatabroker.com/data/?utm_source=huggingface&utm_medium=model&utm_campaign=record-type-page-classifier)
+- [Can you license CRM history to AI?](https://aitrainingdatabroker.com/data/crm-history/?utm_source=huggingface&utm_medium=model&utm_campaign=record-type-page-classifier)
+- [Jira data export for AI licensing](https://aitrainingdatabroker.com/data/jira-data-export/?utm_source=huggingface&utm_medium=model&utm_campaign=record-type-page-classifier)
+
+## License and credit
+
+Apache 2.0. Please credit **AI Training Data Broker** and link to https://aitrainingdatabroker.com/ when you share or build on it.
+
+## About
+
+AI Training Data Broker is a broker and technical advisor for companies that sell or license their data to AI labs. It guides the technical side, including export, format, de-identification and compliance, and publishes free guides, checklists and tools.