2796gauravc's picture
Upload README.md with huggingface_hub
1a1cca7 verified
|
Raw
History Blame Contribute Delete
1.59 kB
---
language: [en, hi]
license: gemma
tags:
- function-calling
- gemma
- kosha
- android
- on-device
- litert
- mediapipe
- finance
base_model: 2796gauravc/kosha-functiongemma-phase0
---
# KOSHA Phase 0 — Android .task File
**KOSHA (कोश)** — Private on-device Indian finance tracker.
This is the **MediaPipe-ready `.task` file** for Android LiteRT deployment.
The fine-tuned merged model is at [2796gauravc/kosha-functiongemma-phase0](https://huggingface.co/2796gauravc/kosha-functiongemma-phase0).
## File Info
| Property | Value |
|---|---|
| Base model | FunctionGemma 270M |
| Quantization | dynamic_int8 |
| KV cache max len | 1024 |
| Size | ~290 MB |
| Format | MediaPipe `.task` (LiteRT) |
## Functions
- `log_expense` — Indian bills, UPI, groceries, fuel, dining
- `log_income` — Salary, freelance, UPI received
- `no_expense` — OTP, promotions, non-financial messages
## Android Integration
```kotlin
// build.gradle.kts
implementation("com.google.mediapipe:tasks-genai:0.10.22")
// Usage
val options = LlmInference.LlmInferenceOptions.builder()
.setModelPath("/data/local/tmp/kosha_phase0_q8_ekv1024.task")
.setMaxTokens(1024)
.setTopK(64)
.setTopP(0.95f)
.setTemperature(0.1f) // Low temp for deterministic function calls
.setPreferredBackend(LlmInference.Backend.CPU) // CPU recommended for Gemma 270M
.build()
val llm = LlmInference.createFromOptions(context, options)
val result = llm.generateResponse(yourFunctionGemmaPrompt)
// Parse: <start_function_call>call:log_expense{...}<end_function_call>
```