BenFrodom/rtila-assistant-lite-1.5

🤖 RTILA Assistant Lite 1.5

One model to rule them all — fine-tuned on Qwen3.5-9B, the next-gen hybrid architecture with Gated Delta Networks

📋 Model Description

RTILA Assistant Lite 1.5 is a unified replacement for the entire previous RTILA model family (Mini, Lite, and full Assistant). Built on Alibaba's Qwen3.5-9B — a fundamentally new architecture featuring Gated Delta Networks and hybrid attention — it delivers flagship-level quality in a single, efficient package that runs on a wide range of hardware.

🔄 Replaces All Previous Versions

Previous Model	Base	GGUF Size	Status
RTILA Assistant	Qwen3-14B	9 GB	❌ Superseded
RTILA Assistant Lite	Qwen3-8B	5 GB	❌ Superseded
RTILA Assistant Mini	Qwen3-4B	2.5 GB	❌ Superseded
RTILA Assistant Lite 1.5	Qwen3.5-9B	~6 GB	✅ Current

✨ Why One Model?

Qwen3.5-9B's hybrid Gated Delta Net + sparse MoE architecture is so efficient that a single 9B model now matches or exceeds the quality of the old 14B model while running in a fraction of the memory. There's no longer a reason to maintain three separate variants.

Metric	Old Full (Qwen3-14B)	Old Lite (Qwen3-8B)	Lite 1.5 (Qwen3.5-9B)
GGUF Size	~9 GB	~5 GB	~6 GB
Min RAM	16 GB	8 GB	8 GB
Architecture	Standard Transformer	Standard Transformer	Hybrid Gated DeltaNet
Base Generation	Qwen3	Qwen3	Qwen3.5
Quality	⭐⭐⭐⭐⭐	⭐⭐⭐⭐	⭐⭐⭐⭐⭐

Capabilities

Category	Description
🌐 Navigation & Interaction	Click, scroll, type, wait, handle popups, multi-tab workflows
📊 Data Extraction	CSS/XPath selectors, tables, lists, nested data, pagination
🔄 Logic & Flow	Loops, conditionals, error handling, retry patterns
🔗 Triggers & Integrations	Webhooks, PostgreSQL, MySQL, Slack, email notifications
📝 Variables & Substitution	Dynamic values, data transformations, regex patterns
🛠️ Advanced Scripting	Custom JavaScript execution, page analysis, DOM manipulation

---

📦 Model Specifications

Property	Value
Base Model	Qwen3.5-9B
Architecture	Hybrid Gated DeltaNet + Gated Attention
Parameters	9B
Format	GGUF Q4_K_M
Size	~6 GB
Context Length	2048 tokens (fine-tuned)

What's New in the Base Model

Qwen3.5-9B is not just an incremental update — it's a new architecture:

• Gated Delta Networks: A hybrid layout of linear attention (DeltaNet) and standard attention layers for high-throughput inference with lower latency
• Unified Vision-Language Foundation: Early fusion training on multimodal tokens
• 201 Language Support: Massively expanded multilingual coverage
• Native 262K Context: The base model supports up to 262,144 tokens natively

---

💻 Hardware Requirements

Hardware	Supported	Notes
GPU (8GB+ VRAM)	✅ Recommended	RTX 3060, RTX 4060, RTX 3070
GPU (6GB VRAM)	⚠️ Tight	May need CPU offloading for some layers
Apple Silicon 16GB+	✅ Excellent	M1/M2/M3/M4 Pro/Max — fast Metal inference
Apple Silicon 8GB	⚠️ Workable	Runs but memory-constrained; close other apps
CPU-only (8GB+ RAM)	✅ Viable	Reasonable inference speed

---

🚀 Quick Start

Option 1: Ollama (Easiest)

# Run directly from Hugging Face
ollama run hf.co/rtila-corporation/rtila-assistant-lite-1.5:Q4_K_M

Or create a custom Modelfile:

FROM hf.co/rtila-corporation/rtila-assistant-lite-1.5:Q4_K_M

PARAMETER temperature 0.7
PARAMETER top_p 0.8
PARAMETER top_k 20

SYSTEM """
You are RTILA Assistant, an expert AI for generating automation configurations for the RTILA Automation Engine.
"""

ollama create rtila-1.5 -f Modelfile
ollama run rtila-1.5

Option 2: LM Studio

1. Download LM Studio
2. Search for rtila-corporation/rtila-assistant-lite-1.5
3. Download Q4_K_M (~6 GB)
4. Set parameters: Temperature=0.7, Top-P=0.8, Top-K=20
5. Start chatting!

Option 3: llama.cpp

# Download model
huggingface-cli download rtila-corporation/rtila-assistant-lite-1.5 \
  qwen3.5-9b-Q4_K_M.gguf --local-dir ./models

# Run interactive chat
./llama-cli -m ./models/qwen3.5-9b-Q4_K_M.gguf \
  -p "Scrape product prices from an e-commerce site" \
  --ctx-size 2048 --temp 0.7 --top-p 0.8 --top-k 20

Option 4: Python (llama-cpp-python)

from llama_cpp import Llama

llm = Llama(
    model_path="./models/qwen3.5-9b-Q4_K_M.gguf",
    n_ctx=2048,
    n_gpu_layers=-1,  # Use GPU/Metal if available, 0 for CPU only
)

response = llm.create_chat_completion(
    messages=[
        {"role": "system", "content": "You are RTILA Assistant..."},
        {"role": "user", "content": "Create a config for scraping product data"}
    ],
    temperature=0.7,
    top_p=0.8,
    top_k=20,
)
print(response['choices'][0]['message']['content'])

---

⚙️ Recommended Generation Settings

Based on Qwen3.5's official recommendations for non-thinking (instruct) mode:

Parameter	Value
Temperature	0.7
Top-P	0.8
Top-K	20
Min-P	0

ℹ️ Note: This model was trained with thinking mode disabled for efficient structured JSON output. It will not produce <think>...</think> blocks.

---

💬 Example Prompts & Outputs

Example 1: Basic Extraction

Prompt:

Extract all product names and prices from the current page

Output:

{
  "settings": {
    "name": "Product Extraction"
  },
  "datasets": {
    "products": {
      "selector": ".product-item",
      "type": "list",
      "fields": {
        "name": {
          "selector": ".product-title",
          "type": "text"
        },
        "price": {
          "selector": ".product-price",
          "type": "text",
          "transform": "extractNumber"
        }
      }
    }
  },
  "commands": [
    {"action": "extract", "dataset": "products"}
  ]
}

Example 2: Pagination Loop

Prompt:

Scrape reviews from multiple pages, clicking "Next" until no more pages

Output:

{
  "settings": {
    "name": "Review Scraper with Pagination"
  },
  "datasets": {
    "reviews": {
      "selector": ".review-card",
      "type": "list",
      "fields": {
        "author": {"selector": ".reviewer-name", "type": "text"},
        "rating": {"selector": ".star-rating", "type": "attribute", "attribute": "data-rating"},
        "text": {"selector": ".review-body", "type": "text"},
        "date": {"selector": ".review-date", "type": "text"}
      }
    }
  },
  "commands": [
    {
      "action": "loop",
      "while": {"exists": ".next-page:not(.disabled)"},
      "maxIterations": 50,
      "commands": [
        {"action": "extract", "dataset": "reviews", "append": true},
        {"action": "click", "selector": ".next-page"},
        {"action": "wait", "duration": 2000}
      ]
    }
  ]
}

Example 3: Database Integration

Prompt:

After extraction, save data to PostgreSQL and send Slack notification

Output:

{
  "triggers": {
    "onComplete": [
      {
        "type": "database",
        "config": {
          "driver": "postgresql",
          "connection": "{{env.DATABASE_URL}}",
          "table": "scraped_data",
          "mode": "upsert",
          "uniqueKey": "id"
        }
      },
      {
        "type": "webhook",
        "config": {
          "url": "{{env.SLACK_WEBHOOK}}",
          "method": "POST",
          "body": {
            "text": "✅ Extraction complete: {{datasets.products.length}} items saved"
          }
        }
      }
    ]
  }
}

Example 4: Knowledge Question

Prompt:

What is the difference between execute_script and run_script?

Output:

execute_script and run_script serve different purposes in RTILA:

**execute_script**: Runs JavaScript synchronously in the page context. Returns immediately 
with the result. Best for quick DOM queries, getting values, or simple manipulations.

**run_script**: Executes JavaScript asynchronously with full control flow. Supports 
await, can make fetch requests, and handles complex async operations. Returns a Promise.

Use execute_script for: Reading values, checking conditions, simple DOM changes
Use run_script for: API calls, complex async workflows, operations that need to wait

---

🏋️ Training Details

Parameter	Value
Base Model	unsloth/Qwen3.5-9B
Method	QLoRA (4-bit)
LoRA Rank	64
LoRA Alpha	128
LoRA Dropout	0.0 (Unsloth optimized)
Target Modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Context Length	2048 tokens
Training Examples	~400
Epochs	6 (with early stopping, patience=4)
Learning Rate	2e-4 (cosine schedule)
Effective Batch Size	16 (batch=2 × grad_accum=8)
Optimizer	AdamW 8-bit
Warmup	10% of total steps
Thinking Mode	Disabled
rsLoRA	✅ Enabled

Training Hardware

Component	Spec
GPU	NVIDIA RTX 3060 12GB
RAM	64 GB
CPU	Intel i7-8700
OS	Windows

Training Data

• Navigation & interaction patterns
• Data extraction configurations
• Logic & flow control
• Triggers & integrations
• Variables & substitution
• Advanced scripting
• Error handling
• Knowledge base Q&A

---

📝 System Prompt

For best results, use this system prompt:

You are RTILA Assistant, an expert AI for generating automation configurations for the RTILA Automation Engine.

Your capabilities:
1. Generate complete JSON configurations for web automation tasks
2. Define datasets with selectors, properties, and transformations
3. Configure navigation, extraction, loops, and conditionals
4. Set up triggers for webhooks, databases, and integrations
5. Explain RTILA concepts and best practices

When generating configurations:
- Always output valid JSON with proper structure
- Include 'settings', 'datasets', and 'commands' sections as needed
- Use appropriate selectors (CSS, XPath) for the target elements
- Apply transformations when data cleaning is required

When answering questions:
- Be concise and accurate
- Provide examples when helpful
- Reference specific RTILA features and commands

---

🔗 Links

Resource	Link
This Model	huggingface.co/rtila-corporation/rtila-assistant-lite-1.5
Base Model	huggingface.co/Qwen/Qwen3.5-9B
RTILA Platform	rtila.com

Previous Models (Superseded)

Model	Status
RTILA Assistant (Qwen3-14B)	Superseded by Lite 1.5
RTILA Assistant Lite (Qwen3-8B)	Superseded by Lite 1.5
RTILA Assistant Mini (Qwen3-4B)	Superseded by Lite 1.5

---

📄 License

Apache 2.0

---

🙏 Acknowledgments

• Built on Qwen3.5-9B by Alibaba
• Fine-tuned with Unsloth
• Quantized with llama.cpp