| --- |
| library_name: transformers |
| tags: |
| - coding |
| - reasoning |
| - language |
| - python |
| - javascript |
| - sql |
| - agentic |
| - lora |
| - sft |
| - qwen |
| license: mit |
| base_model: Qwen/Qwen3.5-4B-Thinking |
| language: |
| - en |
| pipeline_tag: text-generation |
| --- |
| |
| # aiAI Coder V2 |
|
|
| **4B Parameters • Production-Grade Code Generation • Full-Stack System Design** |
|
|
| --- |
|
|
| ## Model Details |
|
|
| - **Base Model:** Qwen/Qwen3.5-4B-Thinking |
| - **Parameter Count:** 4B |
| - **Language:** English |
| - **License:** MIT |
|
|
| --- |
|
|
| ## Model Description |
|
|
| aiAI Coder V2 is a fine-tuned version of Qwen3.5-4B-Thinking, trained on a custom dataset distilled from Grok. This release represents a significant step forward from V1. |
|
|
| V1 (~5.6K examples) produced competent interview-style code. V2 (~28K examples) generates production-grade code across Python, JavaScript, and SQL, and can design scalable, load-balanced applications with proper prompting. |
|
|
| This model was trained on a single cloud 5090 in under 2 hours, keeping costs low while achieving major capability gains. |
|
|
| For user convenience, we have merged the LoRA adapter with the base model so you can download and test directly with Transformers without needing to load the base model separately. |
|
|
| --- |
|
|
| ## Milestone Status |
|
|
| - Code Generation ✅ Achieved |
| - Debugging ✅ Achieved |
| - System Design ✅ Achieved |
| - Production Readiness ✅ Achieved |
| - Scaling Thinking ✅ Achieved |
| - Security Awareness ✅ Achieved |
| - Full-Stack Knowledge ✅ Achieved |
| - Agentic Capability 🚧 In Progress (V3) |
|
|
| ## Reports |
| In a single prompt, aiAI Coder V2 (4B) generated a complete production-ready URL shortener including: |
| PostgreSQL schema with UUID PK, proper indexes, and atomic click counting |
| FastAPI backend with Redis caching, rate limiting, and collision handling |
| Multi-stage Docker + docker-compose with health checks |
| Frontend + scaling strategy for 100M requests/day |
| Security considerations (rate limiting, input validation, etc.) |
| DeepSeek evaluation: 10/10 across Database Schema, Backend API, Caching, Deployment, Scaling, and Security. |
| |
| --- |
|
|
| ## Uses |
|
|
| ### Direct Use |
|
|
| This model is intended for: |
|
|
| - Code generation and completion |
| - Debugging and bug fixing |
| - System design and architecture planning |
| - SQL query generation and optimization |
| - Educational purposes and prototyping |
|
|
| ### Out-of-Scope Use |
|
|
| - Production deployment without human review |
| - Safety-critical systems without validation |
| - Generating malicious code |
| - Any use violating applicable laws |
|
|
| --- |
|
|
| ## Bias, Risks, and Limitations |
|
|
| - May occasionally hallucinate or produce incorrect code. |
| - Generated code should be reviewed for security vulnerabilities. |
| - Primarily trained on English data; other languages may perform poorly. |
| - Context window is large (262K tokens) but may degrade at extreme lengths. |
| - Not fully agentic; best used with tooling wrappers like OpenCode. |
|
|
| ### Recommendations |
|
|
| - Review all generated code before deployment. |
| - Run generated code in sandboxed environments. |
| - Use safety filters for disallowed content. |
|
|
| --- |
|
|
| ## How to Get Started with the Model |
|
|
| ### Load with Transformers |
|
|
| ```python |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
| import torch |
| |
| model_name = "wefamm/aiAI_coder_V2_4B" |
| |
| tokenizer = AutoTokenizer.from_pretrained(model_name) |
| model = AutoModelForCausalLM.from_pretrained( |
| model_name, |
| torch_dtype=torch.bfloat16, |
| device_map="auto" |
| ) |
| |
| messages = [ |
| {"role": "system", "content": "You are a helpful coding assistant."}, |
| {"role": "user", "content": "Write a Python function to reverse a linked list in-place."} |
| ] |
| |
| inputs = tokenizer.apply_chat_template( |
| messages, |
| add_generation_prompt=True, |
| return_tensors="pt" |
| ).to(model.device) |
| |
| outputs = model.generate( |
| inputs, |
| max_new_tokens=1024, |
| temperature=0.2, |
| do_sample=True |
| ) |
| |
| response = tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True) |
| print(response) |
| ``` |
|
|
| Load with vLLM (Production) |
|
|
| ```bash |
| vllm serve wefamm/aiAI_coder_V2_4B \ |
| --max-model-len 8192 \ |
| --tensor-parallel-size 1 \ |
| --dtype bfloat16 |
| ``` |
|
|
| Load with Ollama |
|
|
| Create a Modelfile: |
|
|
| ```dockerfile |
| FROM wefamm/aiAI_coder_V2_4B |
| |
| PARAMETER temperature 0.2 |
| PARAMETER num_ctx 8192 |
| |
| TEMPLATE """{{ if .System }}<|im_start|>system |
| {{ .System }}<|im_end|> |
| {{ end }}{{ if .Prompt }}<|im_start|>user |
| {{ .Prompt }}<|im_end|> |
| {{ end }}<|im_start|>assistant |
| """ |
| ``` |
|
|
| Then run: |
|
|
| ```bash |
| ollama create aiAI-coder -f Modelfile |
| ollama run aiAI-coder |
| ``` |
|
|
| --- |
|
|
| Training Details |
|
|
| Training Data |
|
|
| · Source: Distilled completions from Grok |
| · Size: ~28,000 examples |
| · Format: Multi-turn conversations with reasoning blocks |
| · Focus Areas: Python, JavaScript, SQL, debugging, system design |
|
|
| Training Procedure |
|
|
| · Method: LoRA Supervised Fine-Tuning (SFT) |
| · Epochs: 1 |
| · Hardware: Single NVIDIA 5090 (32GB VRAM) |
| · Training Time: <2 hours |
|
|
|
|
|
|
| --- |
|
|
| Evaluation |
|
|
| Benchmarks |
|
|
| · HumanEval (Pass@1): Coming Soon |
| · LiveCodeBench-v6: 54.2% (base model score) |
|
|
| Internal Testing |
|
|
| The model passed comprehensive custom tests across: |
|
|
| · Advanced algorithms (Manacher's O(n) palindrome) |
| · System design (URL shortener with scaling) |
| · Debugging (identifying subtle bugs with explanations) |
| · OOP (encapsulation, validation) |
| · Async (concurrent downloads) |
| · Database design (PostgreSQL schema) |
| · Security (rate limiting, injection prevention) |
|
|
| --- |
|
|
| Environmental Impact |
|
|
| · Hardware Type: NVIDIA 5090 (32GB VRAM) |
| · Hours Used: <2 hours |
| · Cloud Provider: AutoDL |
| · Carbon Emitted: Estimated ~0.5–1.0 kg CO2 equivalent |
|
|
| --- |
|
|
| Technical Specifications |
|
|
| Hardware |
|
|
| · NVIDIA 5090 with 32GB VRAM |
|
|
| Software |
|
|
| · Transformers |
| · PEFT (LoRA) |
| · PyTorch |
|
|
| --- |
|
|
| Citation |
|
|
| If you use this model in your research or applications, please cite: |
|
|
| ```bibtex |
| @misc{aiAI-coder-V2, |
| author = {aiAI}, |
| title = {aiAI Coder V2: Production-Grade 4B Coding Assistant}, |
| year = {2026}, |
| publisher = {Hugging Face}, |
| howpublished = {\url{https://huggingface.co/aiAI_coder_V2_4B}} |
| } |
| ``` |
|
|
| --- |
|
|
| More Information |
|
|
| V3 is in development and will focus on: |
|
|
| · Full agentic capabilities |
| · Multi-turn task completion |
| · Tool calling integration |
| · Enhanced reasoning |
| · Expanded language support |
|
|
| --- |
|
|
| Model Card Authors |
|
|
| · aiAI |
| · nitrous-0xide (funding) |
|
|
| Model Card Contact |
|
|
| wefam67@proton.me |
|
|
| --- |
|
|
| Acknowledgments |
|
|
| · Base Model: Qwen/Qwen3.5-4B-Thinking by Alibaba |
| · Distillation Source: Grok |
| · Training Infrastructure: AutoDL |
|
|
| --- |
|
|
| This model is imperfect — V3 will be better. We're iterating fast. Expect improvements in the coming days/weeks. |
|
|
| Happy coding! 🚀 |
|
|
