Add README

README.md

@@ -1,105 +1,24 @@
 ---
-language:
-- en
-license: apache-2.0
-library_name: transformers
-base_model: unsloth/Qwen2.5-1.5B-Instruct
 tags:
-- unsloth
-- qwen-2.5
-- mobile-agent
-- tool-use
 - gguf
-- qlora
-datasets:
-- mlabonne/FineTome-100k
-- google/mobile-actions
-model_name: Jotex
----
-
-# Jotex
-
-**Jotex** is a lightweight, 1.7B parameter instruction-tuned and device-action-oriented language model developed by **KielTech**. Built using QLoRA optimization via Unsloth, Jotex is specifically designed to power offline mobile assistants, processing natural language instructions and mapping them to tool calls and local phone-control functions.
-
-## Model Details
-
-* **Developed by:** KielTech
-* **Model Type:** Causal Language Model / Mobile Automation Agent
-* **Base Model:** [Qwen2.5-1.5B-Instruct](https://huggingface.co/unsloth/Qwen2.5-1.5B-Instruct)
-* **Language(s):** English
-* **License:** Apache-2.0
-* **Finetuned from model:** unsloth/Qwen2.5-1.5B-Instruct
-
-### Model Sources
-* **Repository:** [kiel2/Jotex](https://huggingface.co/kiel2/Jotex)
-
----
-
-## Uses
-
-### Direct Use
-* **Offline Mobile Control:** Acting as an on-device local assistant capable of interpreting phone commands and formatting them into structured OS intents or system actions.
-* **General Conversational Reasoning:** Handling multi-turn dialogues, logic tasks, and general queries while retaining low memory overhead.
-
-### Out-of-Scope Use
-* High-stakes enterprise automation without human-in-the-loop validation.
-* Critical infrastructure control or medical diagnoses.
-
----
-
-## Bias, Risks, and Limitations
-* **Hallucination:** Like all models of this scale (~1.7B parameters), Jotex may occasionally generate incorrect data or fail complex multi-step logical inferences.
-* **Scope Boundaries:** The model is optimized for mobile task delegation and conversational reasoning; it is not a general-purpose substitute for ultra-large foundation models (e.g., 70B+ parameters).
-
-### Recommendations
-Users should deploy Jotex within sandboxed device environments where tool-execution logic validates user intents before executing sensitive local device APIs.
-
----
-
-## Training Details
-
-### Training Data
-Jotex was fine-tuned on a balanced, mixed dataset consisting of:
-1. **General Conversational Data:** 5,000 curated rows from [`mlabonne/FineTome-100k`](https://huggingface.co/datasets/mlabonne/FineTome-100k) to preserve linguistic capability and reasoning flow.
-2. **Mobile Function-Calling Data:** Structured rows from [`google/mobile-actions`](https://huggingface.co/datasets/google/mobile-actions) focused on parsing human phrasing into device utility commands.
-
-### Training Hyperparameters
-* **Quantization:** 4-bit LoRA (QLoRA via Unsloth)
-* **Max Sequence Length:** 2048 tokens
-* **Per-Device Batch Size:** 2
-* **Gradient Accumulation Steps:** 4
-* **Effective Batch Size:** 8
-* **Learning Rate:** 2e-4
-* **Warmup Steps:** 5
-* **Training Regime:** FP16 / BF16 mixed precision
+- llama.cpp
+- unsloth
 
 ---
 
-## How to Get Started with the Model
-
-You can load the model locally using Python with Unsloth or Transformers, or use the exported GGUF files (`q4_k_m`) directly within mobile wrappers like `llama.cpp`.
+# Jotex : GGUF
 
-```python
-from unsloth import FastLanguageModel
-import torch
+This model was finetuned and converted to GGUF format using [Unsloth](https://github.com/unslothai/unsloth).
 
-max_seq_length = 2048
-model, tokenizer = FastLanguageModel.from_pretrained(
-    model_name = "kiel2/Jotex",
-    max_seq_length = max_seq_length,
-    load_in_4bit = True,
-)
-FastLanguageModel.for_inference(model) # Enable native fast inference
+**Example usage**:
+- For text only LLMs:    `llama-cli -hf kiel2/Jotex --jinja`
+- For multimodal models: `llama-mtmd-cli -hf kiel2/Jotex --jinja`
 
-# Example prompt structure matching Qwen 2.5 chat template
-messages = [
-    {"role": "user", "content": "Turn on my Wi-Fi and open settings."}
-]
-input_ids = tokenizer.apply_chat_template(
-    messages, 
-    add_generation_prompt=True, 
-    return_tensors="pt"
-.to("cuda"))
+## Available Model files:
+- `qwen2.5-1.5b-instruct.Q8_0.gguf`
+- `qwen2.5-1.5b-instruct.Q4_K_M.gguf`
 
-outputs = model.generate(input_ids=input_ids, max_new_tokens=128)
-print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+## Ollama
+An Ollama Modelfile is included for easy deployment.
+This was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth)
+[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)