coda-llm

Coda LLM (Granite-4.2-8B Najdi Sales Agent)

Coda LLM is an enterprise-grade Saudi sales and voice-agent model built on top of ibm-granite/granite-4.2-8b. It is specifically fine-tuned for high-velocity real-time phone calls, tele-sales, and customer relationship interactions in Saudi Arabia.

Key Capabilities & Pillars

1. Native Najdi Dialect with Full Tashkeel (TTS-Grade):

   • Natural Saudi business and everyday Najdi vernacular ("أبشر", "يا هلا والله", "وش الباقة اللي تناسبك؟", "تمام كذا").
   • Diacritized outputs optimized directly for TTS (Text-to-Speech) engines with clean phonetic clarity and zero MSA robotic leaks.

2. Native Function Calling & Tool Use:

   • Evaluated across real-world enterprise sales tools (search_products, calculate_quote, schedule_appointment, create_subscription, check_order_status).
   • 100% negative rejection rate on irrelevant queries (e.g. general chit-chat, weather) — never hallucinates unrequested tool calls.
   • 100% grounded pricing: strictly extracts price quotes from tool outputs instead of inventing figures.

3. Autonomous Sales Intelligence:

   • Discovery & Qualification: Actively asks clarifying questions about needs, budget, and family/team size before proposing packages.
   • Objection Handling: Handles price objections via value framing and flexible plans instead of discount dumping.
   • Immediate Stop-Selling: Strictly respects customer rejection cues ("ماني مهتم", "لا تتصلون علي") with 100% immediate graceful exit.

4. Reasoning & Deliberation:

   • Generates internal thinking thoughts (<think>...</think>) before tool calling or conversational delivery to ensure multi-step sales constraints are honored.

---

Benchmark Results (v10 Production Release)

Evaluated across a comprehensive 4-pillar, 27-scenario multi-trial suite:

---

Quickstart & Usage

1. Serving with vLLM (Recommended for Production & Voice Pipelines)

Run the OpenAI-compatible vLLM server:

python -m vllm.entrypoints.openai.api_server \
    --model mohameddalii/coda-llm \
    --served-model-name coda-llm \
    --tool-call-parser qwen3_coder \
    --reasoning-parser deepseek_r1 \
    --enable-auto-tool-choice \
    --port 8000

Note on Speculative Decoding: Empirical load-testing demonstrates that n-gram speculative decoding offers no speedup on diacritized Arabic text due to novel morphological generation. Serve in standard mode for optimal throughput and low latency.

---

2. Multi-turn Python Client (with Function Calling)

import json
import requests

API_URL = "http://localhost:8000/v1/chat/completions"

tools = [
    {
        "type": "function",
        "function": {
            "name": "search_products",
            "description": "Search available telecom and internet packages.",
            "parameters": {
                "type": "object",
                "properties": {
                    "category": {"type": "string", "description": "e.g. fiber, 5g, business"}
                },
                "required": ["category"]
            }
        }
    },
    {
        "type": "function",
        "function": {
            "name": "calculate_quote",
            "description": "Calculate exact subscription cost.",
            "parameters": {
                "type": "object",
                "properties": {
                    "product_id": {"type": "string"},
                    "quantity": {"type": "integer"}
                },
                "required": ["product_id"]
            }
        }
    }
]

messages = [
    {"role": "system", "content": "أنت ممثل مبيعات اتصالات بلهجة نجدية طبيعية مع التشكيل الكامل."},
    {"role": "user", "content": "وش عندكم عروض فايبر للمنزل؟"}
]

payload = {
    "model": "coda-llm",
    "messages": messages,
    "tools": tools,
    "temperature": 0.7,
    "max_tokens": 1024
}

response = requests.post(API_URL, json=payload).json()
assistant_message = response["choices"][0]["message"]
print("Assistant Response:", assistant_message)

---

3. Loading with Hugging Face Transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "mohameddalii/coda-llm"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

messages = [
    {"role": "system", "content": "أنت ممثل مبيعات سعودي بلهجة نجدية طبيعية مع التشكيل الكامل."},
    {"role": "user", "content": "السلام عليكم، وش أرخص باقة عندكم؟"}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

---

Training Details

• Base Model: ibm-granite/granite-4.2-8b
• Method: QLoRA Fine-Tuning merged directly into full 16-bit safetensors weights.
• Dataset: Highly curated dataset of real multi-turn Saudi tele-sales conversations, anti-hallucination negative examples, tool calling schemas, and diacritized speech synthesis pairs.
• Context Length: Supports up to 131,072 tokens natively.