---
language:
- en
license: apache-2.0
base_model: Qwen/Qwen2.5-Coder-7B-Instruct
tags:
- lora
- peft
- coding
- agentic
- tool-use
- conductor
- ollama
- qwen2.5
model_name: Lumen-7b-v2
pipeline_tag: text-generation
---

# Lumen 7B v2

<div align="center">

**Lumen** is an agentic AI coding assistant built by [Alexander Wondwossen](https://github.com/thealxlabs) (TheAlxLabs).  
Fine-tuned on Qwen2.5-Coder-7B-Instruct with LoRA for tool-use, git, GitHub, and Conductor integration.

</div>

---

## What is Lumen?

Lumen is a locally-running agentic coding AI designed to work inside [Conductor](https://github.com/thealxlabs/conductor). It can:

- Write, read, and edit code and files
- Run shell commands and verify results
- Use git and GitHub (commits, branches, PRs, Actions, secrets)
- Debug TypeScript, Python, Node.js, and Bash
- Call Conductor plugins as tools
- Control your development environment autonomously

---

## Model Details

| Property | Value |
|---|---|
| **Base Model** | Qwen/Qwen2.5-Coder-7B-Instruct |
| **Fine-tuning Method** | QLoRA (4-bit, NF4) |
| **LoRA Rank** | 32 |
| **LoRA Alpha** | 64 |
| **Training Epochs** | 3 |
| **Max Sequence Length** | 2048 |
| **Parameters** | ~7B |
| **GGUF (Q4_K_M)** | `lumen-q4.gguf` (~4.4GB) |
| **Built by** | Alexander Wondwossen — TheAlxLabs, Toronto, Canada |

---

## Quickstart with Ollama

```bash
ollama pull thealxlabs/lumen
ollama run thealxlabs/lumen "What are you?"
```

---

## Quickstart with Transformers

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

base = "Qwen/Qwen2.5-Coder-7B-Instruct"
adapter = "alxstuff/Lumen-7b-v2"

model = AutoModelForCausalLM.from_pretrained(base, torch_dtype=torch.float16)
model = PeftModel.from_pretrained(model, adapter)
tokenizer = AutoTokenizer.from_pretrained(base)

messages = [
    {"role": "system", "content": "You are Lumen, an agentic AI coding assistant built by Alexander (TheAlxLabs)."},
    {"role": "user", "content": "Create a Python script that fetches weather data."}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

---

## System Prompt

```
You are Lumen, an agentic AI coding assistant built by Alexander (TheAlxLabs). 
You run inside Conductor. You have tools: run_shell, read_file, write_file, conductor_plugin. 
Think step-by-step. Use tools to verify.
```

---

## Tools Lumen Knows

| Tool | Description |
|---|---|
| `run_shell` | Execute terminal commands |
| `read_file` | Read file contents |
| `write_file` | Write or create files |
| `conductor_plugin` | Call any Conductor plugin |

---

## Training Data

Lumen was trained on curated agentic multi-turn conversations covering:

- Git workflows (commit, branch, push, reset, rebase, cherry-pick)
- GitHub (PRs, issues, Actions CI, secrets)
- TypeScript / Node.js debugging
- Python virtual environments and debugging
- Bash scripting and disk management
- Conductor plugin installation and debugging
- Port conflicts and environment variable issues
- Lumen self-knowledge (identity, capabilities)

---

## Hardware Requirements

| Setup | Min RAM | Recommended |
|---|---|---|
| Ollama Q4_K_M | 8GB | 16GB+ |
| Transformers (float16) | 16GB | 24GB+ |
| Training (QLoRA) | 16GB VRAM | 24GB VRAM |

---

## Links

- 🤗 **HuggingFace**: [alxstuff/Lumen-7b-v2](https://huggingface.co/alxstuff/Lumen-7b-v2)
- 🦙 **Ollama**: [thealxlabs/lumen](https://ollama.com/thealxlabs/lumen)
- 🐙 **GitHub**: [thealxlabs](https://github.com/thealxlabs)

---

## License

Apache 2.0 — same as the base model.

---

*Built with ❤️ by Alexander Wondwossen — TheAlxLabs, Toronto, Canada*