Update README.md

README.md

@@ -5,14 +5,25 @@ tags:
 - sentence-transformers
 - text-classification
 pipeline_tag: text-classification
+datasets:
+- mserras/alpaca-es-hackaton
+- somosnlp/somos-clean-alpaca-es
+language:
+- es
 ---
 
 # mserras/setfit-alpaca-es-unprocessable-sample-detection
 
-This is a [SetFit model](https://github.com/huggingface/setfit) that can be used for text classification. The model has been trained using an efficient few-shot learning technique that involves:
+This is a [SetFit model](https://github.com/huggingface/setfit) that can be used for filtering the Alpaca ES instruction dataset. 
+
+ This model has been developed during the 2023 Hackaton organized by [SomosNLP](https://somosnlp.org/)/[HF Card](https://huggingface.co/somosnlp) and with the GPUs provided by [Q Blocks](https://www.qblocks.cloud)
+ 
+This model has been trained over "unprocessable" samples of the translated [Clean Alpaca Es](https://huggingface.co/datasets/somosnlp/somos-clean-alpaca-es) dataset from 
+the HF [Argilla](https://argilla.io) space https://huggingface.co/spaces/mserras/somos-alpaca-es.
+
+To this end, a custom tag is proposed: "unprocessable" which corresponds to instruction/input/output triplets that require processing image, fetching information from the 
+open web and similar tasks where the LLM has no capability action, thus, ending in hallucinations or strange outcomes.
 
-1. Fine-tuning a [Sentence Transformer](https://www.sbert.net) with contrastive learning.
-2. Training a classification head with features from the fine-tuned Sentence Transformer.
 
 ## Usage
 
@@ -26,13 +37,36 @@ You can then run inference as follows:
 
 ```python
 from setfit import SetFitModel
+import argilla as rg
+
 
 # Download from Hub and run inference
 model = SetFitModel.from_pretrained("mserras/setfit-alpaca-es-unprocessable-sample-detection")
-# Run inference
-preds = model(["i loved the spiderman movie!", "pineapple on pizza is the worst 🤮"])
+
+def instruct_fields_to_text(field_instruction: str, field_input: str, field_output: str):
+    """Given the instruction, input and output fields, return a text to be used by setfit"""
+    return f"INSTRUCTION:\n{field_instruction}\nINPUT:\n{field_input}\nOUTPUT:\n{field_output}\n"
+
+def sample_to_text(sample: rg.TextClassificationRecord) -> str:
+    """Converts and Argilla TextClassificationRecord to a text to be used by setfit"""
+    return instruct_fields_to_text(sample.inputs["1-instruction"], sample.inputs["2-input"], sample.inputs["3-output"])
+
+# For a given Argilla record:
+
+unprocessable_score = model.predict_proba([sample_to_text(argilla_record)])[0].tolist()[1]
+
 ```
 
+## Evaluation
+
+*Disclaimer*: There was no formal evaluation done, just a bunch of guys looking at the data & the outcomes.
+
+## Changelog
+
+- [09/04/2023] SQL code generation, date conversion, percentual discounts and renewable energies no longer detected as unprocessable.
+- [06/04/2023] It no longer detects password generation as unprocessable.
+
+
 ## BibTeX entry and citation info
 
 ```bibtex
@@ -46,4 +80,4 @@ publisher = {arXiv},
 year = {2022},
 copyright = {Creative Commons Attribution 4.0 International}
 }
-```
+```