# Qwen3-8B CodeAgent 🤖💻

**A coding & agentic reasoning expert built on Qwen3-8B**

> Expert at coding, step-by-step reasoning, data visualization, tool calling, and research paper analysis

## 🎯 Capabilities

| Capability | How it was trained | Dataset |
|---|---|---|
| **Coding** (any language) | SFT on code instructions + competitions | CodeFeedback + Magicoder + OpenCodeReasoning |
| **Agentic Reasoning** | Chain-of-thought with `<think>` blocks | nvidia/OpenCodeReasoning (R1-style traces) |
| **Data Visualization** | Chart/graph code generation | TIGER-Lab/VisCode-200K |
| **Tool Calling** | Function calling with JSON schemas | glaive-function-calling-v2 |
| **Anti-hallucination** | Step-by-step verification, assistant-only loss masking | All datasets with system prompt enforcement |

## 🏗️ Architecture

- **Base Model**: [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) (8.2B params, Apache 2.0)
- **Fine-tuning**: QLoRA (4-bit NF4, r=64, alpha=16, RSLoRA)
- **Target modules**: all-linear (attention + MLP)
- **Training**: SFT with assistant-only loss masking
- **Context**: 4096 tokens (native 32K, extendable to 131K with YaRN)

## 📊 Training Recipe

Based on research from:
- **Qwen3-Coder-Next** (arxiv: 2603.00729) — agentic coding training pipeline
- **Qwen2.5-Coder** (arxiv: 2409.12186) — coarse-to-fine SFT methodology
- **LoRA Without Regret** — high-rank LoRA with RSLoRA scaling
- **VisCoder** (arxiv: 2506.03930) — visualization code generation
- **FLAME** (arxiv: 2405.01525) — factuality-aware alignment

### Hyperparameters
| Parameter | Value |
|---|---|
| Learning rate | 2e-4 (10× base for LoRA) |
| LR scheduler | Cosine with 5% warmup |
| Epochs | 2 |
| Batch size | 16 (2 × 8 grad accum) |
| Max sequence length | 4096 |
| LoRA rank | 64 |
| LoRA alpha | 16 |
| Weight decay | 0.01 |
| Optimizer | AdamW |
| Precision | BF16 + TF32 |

### Dataset Mix (~50K samples)
| Dataset | Samples | Purpose |
|---|---|---|
| TIGER-Lab/VisCode-200K | 12,000 | Visualization & chart generation |
| m-a-p/CodeFeedback-Filtered-Instruction | 10,000 | Code instruction following |
| nvidia/OpenCodeReasoning | 10,000 | Code reasoning with `<think>` traces |
| glaiveai/glaive-function-calling-v2 | 8,000 | Function/tool calling |
| ise-uiuc/Magicoder-OSS-Instruct-75K | 10,000 | Code generation |

## 🚀 Quick Start

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "sukritvemula/Qwen3-8B-CodeAgent"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")

messages = [
    {"role": "system", "content": "You are an expert coding assistant."},
    {"role": "user", "content": "Write a Python function to visualize a binary tree using matplotlib."}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=2048, temperature=0.6, top_p=0.95, top_k=20)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
```

## 🔧 Inference Speed

| Hardware | Speed (tok/s) | Notes |
|---|---|---|
| A100 80GB (BF16) | ~100-150 | Full precision |
| A10G 24GB (BF16) | ~40-50 | Meets 40 tok/s target |
| RTX 4090 (BF16) | ~60-80 | Consumer GPU |
| Any GPU (AWQ INT4) | 2× above | Minimal quality loss |

**Recommended deployment**: [vLLM](https://github.com/vllm-project/vllm) or [SGLang](https://github.com/sgl-project/sglang)

```bash
# vLLM
vllm serve sukritvemula/Qwen3-8B-CodeAgent --enable-reasoning --reasoning-parser deepseek_r1

# SGLang
python -m sglang.launch_server --model-path sukritvemula/Qwen3-8B-CodeAgent --reasoning-parser qwen3
```

## 📝 Training Script

See `train_coding_agent.py` in this repo for the full training pipeline.

## 🗺️ Roadmap (Next Steps)

1. **Stage 2 — GRPO**: Reinforcement learning with code execution reward for improved reasoning
2. **Stage 3 — DPO**: Factuality alignment using FLAME methodology
3. **Multimodal**: Fine-tune Qwen3-VL-7B variant for image understanding + code generation
4. **Scale up**: Increase to 200K+ training samples across all domains

## 📄 License

Apache 2.0 (inherited from Qwen3-8B)