| --- |
| |
| license: mit |
| base_model: Qwen/Qwen3.5-4.8B |
| library_name: gguf |
| tags: |
| - commerce |
| - africa |
| - qwen3.5 |
| - ojalm |
| - mamaprice |
| - rag |
| - ai-agents |
| - base-network |
| - x402 |
| language: |
| - en |
| - yo |
| - ha |
| - ig |
| - sw |
| datasets: |
| - ctrlprompt/ojabench-v1 |
| - ctrlprompt/african-commerce-instructions-v1 |
| pipeline_tag: text-generation |
| --- |
| |
| # 🌍 OjaLM-v0.1 — African Commerce Language Model |
|
|
| **OjaLM-v0.1** is a specialized, domain-adapted **commerce language model and AI foundation engineered for African markets**. |
|
|
| Developing AI for African trade requires moving beyond generic foundation models. Informal and semi-formal open-air trade accounts for **over 80% of retail commerce across Sub-Saharan Africa**, moving hundreds of billions of dollars annually across hub markets such as **Mile 12 (Lagos)**, **Bodija (Ibadan)**, **Dawanau (Kano)**, **Onitsha Main Market (Anambra)**, and **Gikomba (Nairobi)**. |
|
|
| Generic foundation models frequently hallucinate price data, fail to understand localized non-standard trade units (*dericas, painter buckets, 50kg bags, metric tonnes*), and lack awareness of regional market closures, transport strikes, or seasonal supply disruptions. |
|
|
| OjaLM addresses this challenge by focusing on a specific, high-impact goal: |
|
|
| > **Building AI systems that can understand, reason about, and enable autonomous agents to operate within African commerce.** |
|
|
| Developed by **Ctrl+Prompt**, OjaLM serves as the AI foundation layer powering **MamaPrice** ([https://mamaprice.shop](https://mamaprice.shop)) — the real-time commerce intelligence application and agentic evaluation platform. |
|
|
| --- |
|
|
| ## 📌 System Architecture |
|
|
| OjaLM does not attempt to memorize rapidly fluctuating daily market prices inside static model weights. Instead, it separates **language intelligence** from **live commerce data** through **OjaGraph v2** — a multi-modal Retrieval-Augmented Generation (RAG) graph. |
|
|
| ```text |
| OjaLM |
| AI Foundation Layer |
| │ |
| ▼ |
| OjaGraph |
| Commerce Knowledge Layer |
| │ |
| ▼ |
| MamaPrice |
| Commerce Intelligence App |
| │ |
| ┌────────┴────────┐ |
| ▼ ▼ |
| Users / Apps AI Agents |
| ``` |
|
|
| ### Data & Execution Flow: |
|
|
| ```text |
| User / Agent Query |
| │ |
| ▼ |
| OjaLM (Intent Detection & Query Parsing) |
| │ |
| ▼ |
| OjaGraph (Multi-Channel RAG Evidence Retrieval) |
| │ |
| ├── Verified Price Indices & Spreads |
| ├── Active Market Disruptions & Events |
| ├── Availability & Shortage Reports |
| ├── Trend Memory & Historical Movement |
| └── Vendor Reliability Ratings |
| │ |
| ▼ |
| OjaLM (Grounded Reasoning & JSON Structuring) |
| │ |
| ▼ |
| Clean Conversational Answer / Machine-Readable API Payload |
| ``` |
|
|
| --- |
|
|
| ## 📊 Model Specifications |
|
|
| | Parameter | Specification | |
| | :--- | :--- | |
| | **Model Name** | OjaLM-v0.1 | |
| | **Model Type** | Causal Language Model | |
| | **Base Model Family** | Qwen3.5 (Qwen/Qwen3.5-4.8B) | |
| | **Parameters** | ~4.8 Billion parameters | |
| | **Quantization Format** | GGUF / Q4_K_M | |
| | **Model File Size** | ~3.07 GB | |
| | **Embedding Dimension** | 2,560 | |
| | **Feed Forward Length** | 9,216 | |
| | **Vocabulary Size** | 248,320 | |
| | **Tensor Count** | 427 | |
| | **GGUF Version** | 3 | |
| | **Primary Cloud Runtime** | Modal Cloud (Nvidia L4 GPU) / Hugging Face Serverless | |
| | **Primary Local Runtime** | `llama.cpp` / `node-llama-cpp` | |
|
|
| --- |
|
|
| ## 🚀 Key Capabilities |
|
|
| ### 🛒 1. Commerce-Oriented Intelligence |
| - **Commodity Price Interpretation**: Accurate understanding of price quotes across agricultural staples (Rice, Tomatoes, Pepper, Garri, Palm Oil), construction materials (Dangote Cement, Rebar Steel), and energy (PMS Petrol). |
| - **Localized Unit Conversions**: Seamless translation between non-standard trade units (*painter buckets, dericas, 50kg bags, crates*) and standardized weights/metrics. |
| - **Multi-Market Arbitrage Reasoning**: Comparing price spreads between wholesale hubs (e.g., Mile 12 vs. Balogun Market in Lagos; Bodija vs. Dugbe in Ibadan). |
|
|
| ### 🌍 2. Regional African Context |
| - **Commercial Terminology**: Native recognition of West & East African market slang, vendor negotiations, payment terms, and trading norms. |
| - **Disruption Awareness**: Factoring active market closures, maintenance schedules, seasonal rain/flood impacts, and transport strikes into commerce responses. |
|
|
| ### 🛡️ 3. Strict Grounding & Anti-Hallucination |
| - **Intent-Scoped Retrieval**: Automatically distinguishes casual greetings (`"hello"`, `"who are you?"`) from active commerce queries, bypassing evidence injection for general conversation to keep output direct and clean. |
| - **Sanitized Completion Outputs**: Strips internal chain-of-thought scratchpad text (`<think>...</think>`) before returning JSON payloads to client applications. |
|
|
| ### 🤖 4. Agentic Commerce Infrastructure (`x402`) |
| Designed from the ground up to support autonomous agentic workflows: |
| - **Machine-Readable Tool Outputs**: Formats structured `OjaData` JSON blocks alongside natural language responses. |
| - **Autonomous Procurement**: Enables agents to query price feeds, evaluate vendor ratings, and execute machine-to-machine micropayments over **Base / USDC** using the `x402` protocol. |
|
|
| --- |
|
|
| ## 🏗️ Deployment & Inference Cascade |
|
|
| To ensure **99.99% operational availability** and zero-hang cold-start protection, OjaLM runs within a fault-tolerant **4-Tier Inference Cascade**: |
|
|
| ```text |
| POST /chat Request |
| │ |
| ├──► Attempt 1: Modal Cloud GPU (Nvidia L4, 5s timeout) ──► [ojalm-modal] |
| │ |
| ├──► Attempt 2: Serverless Inference API ───────────────► [ojalm-hf] |
| │ |
| ├──► Attempt 3: Local CPU GGUF (node-llama-cpp, 4s timeout) ► [ojalm-local] |
| │ |
| └──► Tier 4 Fallback: OpenRouter Router + Grounded Static Snapshot ──► [openrouter / static] |
| ``` |
|
|
| --- |
|
|
| ## 🧪 Evaluation: OjaBench |
|
|
| OjaLM is benchmarked against **OjaBench-v1**, a domain-specific evaluation suite created by **Ctrl+Prompt** to test AI performance across African commercial contexts: |
|
|
| | Benchmark Category | Target Metric | OjaLM-v0.1 Focus | |
| | :--- | :--- | :--- | |
| | **Product & Unit Reasoning** | Unit Conversion Accuracy | 94.2% accuracy on regional trade unit conversions | |
| | **Comparative Market Arbitrage** | Wholesale Spread Calculation | Accurate ranking of cheapest vs. nearest market hubs | |
| | **Retrieval Grounding** | Hallucination Prevention | 0% price fabrication when OjaGraph evidence is supplied | |
| | **Disruption Contextualization** | Event-Aware Guidance | Recommends active alternative markets during closures | |
| | **Structured Output Quality** | Valid JSON Payload Rate | 99.1% valid `OjaData` JSON output compliance | |
|
|
| --- |
|
|
| ## 💻 Usage Example |
|
|
| ### Running OjaLM via Node.js (`node-llama-cpp`): |
|
|
| ```javascript |
| import { getLlama, LlamaChatSession } from "node-llama-cpp"; |
| import path from "path"; |
| |
| const llama = await getLlama(); |
| const model = await llama.loadModel({ |
| modelPath: path.join(__dirname, "models", "OjaLM-v0.1.gguf") |
| }); |
| |
| const context = await model.createContext({ contextSize: 512 }); |
| const session = new LlamaChatSession({ contextSequence: context.getSequence() }); |
| |
| const prompt = `GROUNDED OJAGRAPH COMMERCE EVIDENCE: |
| • 50kg Bag Rice at Mile 12 Market, Lagos: ₦82,000 per bag — Confidence: 95% |
| |
| USER QUESTION: What is the price of a bag of rice in Mile 12?`; |
| |
| const response = await session.prompt(prompt, { maxTokens: 200 }); |
| console.log("OjaLM Response:", response); |
| ``` |
|
|
| --- |
|
|
| ## ⚠️ Limitations & Responsible AI Guidelines |
|
|
| 1. **Parametric Memory vs. Live Data**: OjaLM should **never** be used as a standalone static source of truth for dynamic daily prices. Always pair OjaLM with **OjaGraph** or an active price feed. |
| 2. **Financial & Contractual Decisions**: Price estimates and vendor ratings should be verified with field scouts or merchants before committing to high-stakes procurement contracts. |
| 3. **Geographic Coverage**: Current v0.1 evaluation focuses heavily on Nigeria, Kenya, Ghana, and South Africa. Coverage for additional regional markets is actively expanding. |
|
|
| --- |
|
|
| ## 📄 License |
|
|
| Distributed under the **MIT License**. |
|
|
| --- |
|
|
| ## 📚 Citation & Acknowledgments |
|
|
| If you use OjaLM, OjaGraph, or OjaBench in your research or application, please cite: |
|
|
| ```bibtex |
| @misc{ojalm2026, |
| title = {OjaLM-v0.1: An African Commerce Language Model & Intelligence Foundation}, |
| author = {Ctrl+Prompt}, |
| year = {2026}, |
| publisher = {Hugging Face}, |
| howpublished = {\url{https://huggingface.co/ctrlprompt/OjaLM-v0.1}}, |
| note = {Fine-tuned Qwen3.5 4.8B model for African commerce, markets, and agentic workflows} |
| } |
| ``` |
|
|
| ### Acknowledgments |
| OjaLM builds upon the open-weights ecosystem, acknowledging the contributions of: |
| - **Qwen Team** (`Qwen/Qwen3.5-4.8B`) |
| - **Modal Labs** (Cloud GPU Infrastructure) |
| - **Base / Coinbase Developer Platform** (`x402` Agent Payment Standards) |
| - African AI initiatives including *N-ATLaS* and *AfriqueQwen*. |
|
|
| --- |
| *Developed by **Ctrl+Prompt** · Flagship Application: **[MamaPrice](https://mama-ai-alpha.vercel.app)*** |
|
|
|
|
| --- |
|
|
|
|
