Update README.md

README.md

@@ -1,58 +1,79 @@
----
-base_model: google/functiongemma-270m-it
-library_name: transformers
-model_name: functiongemma-tasky
-tags:
-- generated_from_trainer
-- sft
-- trl
-licence: license
----
+  ---
+  language:
+  - en
+  - it
+  base_model: google/functiongemma-270m-it
+  tags:
+  - function-calling
+  - tool-calling
+  - task-management
+  - todo
+  - synthetic
+  - transformers
+  library_name: transformers
+  pipeline_tag: text-generation
+  ---
 
-# Model Card for functiongemma-tasky
+  # Tasky
 
-This model is a fine-tuned version of [google/functiongemma-270m-it](https://huggingface.co/google/functiongemma-270m-it).
-It has been trained using [TRL](https://github.com/huggingface/trl).
+  ## Model Summary
+  This model is a fine-tuned **function-calling assistant** for a todo/task application. It maps user requests to one of four tools and produces valid tool
+  arguments according to the schema in `AI-TRAINING-TOOLS.md`.
 
-## Quick start
+  - **Base model:** `google/functiongemma-270m-it`
+  - **Primary languages:** English and Italian (with light spelling errors/typos to mimic real users)
+  - **Task:** Structured tool selection + argument generation
 
-```python
-from transformers import pipeline
+  ## Intended Use
+  Use this model to translate natural language task requests into tool calls for:
+  - `create_tasks`
+  - `search_tasks`
+  - `update_tasks`
+  - `delete_tasks`
 
-question = "If you had a time machine, but could only go to the past or the future once and never return, which would you choose and why?"
-generator = pipeline("text-generation", model="independently-platform/functiongemma-tasky", device="cuda")
-output = generator([{"role": "user", "content": question}], max_new_tokens=128, return_full_text=False)[0]
-print(output["generated_text"])
-```
+  It is designed for **task/todo management** workflows and should be paired with strict validation of tool arguments before execution.
 
-## Training procedure
+  ### Example
+  **Input (user):**
 
- 
+  Aggiungi un task per pagare la bolletta della luce domani mattina
 
 
-This model was trained with SFT.
+  **Expected output (model):**
+  ```json
+  {
+    "tool_name": "create_tasks",
+    "tool_arguments": "{\"tasks\":[{\"content\":\"pagare la bolletta della luce\",\"dueDate\":\"2026-01-13T09:00:00.000Z\"}]}"
+  }
 
-### Framework versions
+  ## Training Data
 
-- TRL: 0.26.2
-- Transformers: 4.57.3
-- Pytorch: 2.9.1+cu128
-- Datasets: 4.4.2
-- Tokenizers: 0.22.2
+  Synthetic, bilingual tool-calling data built from the tool schema, including:
 
-## Citations
+  - Multiple phrasings and paraphrases
+  - Mixed English/Italian prompts
+  - Light typos and user mistakes in user_content
+  - Broad coverage of optional parameters
 
+  Splits:
 
+  - Train: 1,500 examples
+  - Eval: 500 examples
 
-Cite TRL as:
-    
-```bibtex
-@misc{vonwerra2022trl,
-	title        = {{TRL: Transformer Reinforcement Learning}},
-	author       = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallou{\'e}dec},
-	year         = 2020,
-	journal      = {GitHub repository},
-	publisher    = {GitHub},
-	howpublished = {\url{https://github.com/huggingface/trl}}
-}
-```
+  ## Training Procedure
+
+  - Fine-tuning on synthetic tool-calling samples
+  - Deduplicated examples
+  - Balanced coverage of all tools and key parameters
+
+  ## Evaluation
+
+  Reported success rate: 99.5% on the 500‑example eval split vs 0% base model.
+  Success was measured as exact match on the predicted tool name and the JSON arguments after normalization.
+
+  ## Limitations
+
+  - Trained for a specific tool schema; not a general-purpose assistant.
+  - Outputs may include incorrect or incomplete tool arguments; validate before execution.
+  - Language coverage is strongest in English and Italian.
+  - Synthetic data may not capture all real-world user phrasing or ambiguity.