File size: 2,247 Bytes
b2343db ef3ae22 244fe7e ef3ae22 b2343db 244fe7e ef3ae22 244fe7e ef3ae22 244fe7e ef3ae22 244fe7e ef3ae22 244fe7e ef3ae22 244fe7e ef3ae22 244fe7e ef3ae22 244fe7e ef3ae22 244fe7e ef3ae22 244fe7e ef3ae22 244fe7e ef3ae22 244fe7e ef3ae22 244fe7e b2343db | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 | ---
datasets:
- independently-platform/tasky
language:
- en
- it
base_model:
- google/functiongemma-270m-it
library_name: transformers
---
# Tasky
## About the model
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`.
- **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
## Intended Use
Use this model to translate natural language task requests into tool calls for:
- `create_tasks`
- `search_tasks`
- `update_tasks`
- `delete_tasks`
It is designed for **task/todo management** workflows and should be paired with strict validation of tool arguments before execution.
### Example
**Input (user):**
Aggiungi un task per pagare la bolletta della luce domani mattina
**Expected output (model):**
```json
{
"tool_name": "create_tasks",
"tool_arguments": "{\"tasks\":[{\"content\":\"pagare la bolletta della luce\",\"dueDate\":\"2026-01-13T09:00:00.000Z\"}]}"
}
## Training Data
Synthetic, bilingual tool-calling data built from the tool schema, including:
- 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
## 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. |