---
license: apache-2.0
base_model: Qwen/Qwen1.5-0.5B
library_name: peft
language:
  - en
tags:
  - lora
  - peft
  - qwen
  - edge-ai
  - edge-impulse
  - documentation
  - code-generation
  - conversational
pipeline_tag: text-generation
---

# edgeai-docs-embedding-qwen1.5-0.5b-instruct

A LoRA adapter fine-tuned on **1,794 Edge Impulse / Edge AI MDX documentation files** from the [Edge Impulse documentation](https://docs.edgeimpulse.com), built on top of [`Qwen/Qwen1.5-0.5B`](https://huggingface.co/Qwen/Qwen1.5-0.5B).

Optimized for:
- Answering developer questions about Edge Impulse Studio, SDK, and APIs
- Summarizing technical documentation
- Generating code snippets for edge ML workflows
- Lightweight local/edge deployment

> **Larger variants in training:** [1.5B](https://huggingface.co/eoinedge/edgeai-qwen2.5coder-1.5b-lora) · [7B](https://huggingface.co/eoinedge/edgeai-qwen2.5coder-7b-lora) (Qwen2.5-Coder base)

---

## Model Details

| Property | Value |
|---|---|
| Base model | `Qwen/Qwen1.5-0.5B` |
| Adapter type | LoRA (PEFT) |
| LoRA rank (`r`) | 8 |
| LoRA alpha | 32 |
| Target modules | `q_proj`, `v_proj` |
| LoRA dropout | 0.05 |
| Trainable parameters | ~786K (0.17% of base) |
| Training epochs | 3 |
| Batch size | 4 (× grad accum 2 = effective 8) |
| Learning rate | 3e-4 (cosine decay) |
| Sequence length | 512 tokens |
| Training hardware | Apple M1 Pro (MPS, fp16) |
| Precision | float16 |

---

## Training Data

| Stat | Value |
|---|---|
| Source | [Edge Impulse Mintlify documentation](https://docs.edgeimpulse.com) |
| File format | MDX (Markdown + JSX components) |
| Total files | 1,794 `.mdx` files |
| Preprocessing | Frontmatter, JSX tags, imports stripped; code fences unwrapped; links flattened |
| Chunks generated | ~3,500 × 512-token chunks |

Topics covered: Studio projects, datasets, DSP blocks, learning blocks, deployment targets, Python SDK, REST API, CLI tools, edge inference, model optimization, and more.

---

## Usage

### Load and chat

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

BASE_MODEL = "Qwen/Qwen1.5-0.5B"
ADAPTER    = "eoinedge/edgeai-docs-embedding-qwen1.5-0.5b-instruct"

device = "cuda" if torch.cuda.is_available() else \
         "mps"  if torch.backends.mps.is_available() else "cpu"

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
base  = AutoModelForCausalLM.from_pretrained(BASE_MODEL,
            dtype=torch.float16 if device != "cpu" else torch.float32,
            device_map=device)
model = PeftModel.from_pretrained(base, ADAPTER)
model.eval()
```

### Single question

```python
prompt = "How do I collect sensor data using the Edge Impulse data forwarder?"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=512,
                         do_sample=True, temperature=0.7, top_p=0.95,
                         pad_token_id=tokenizer.eos_token_id)

print(tokenizer.decode(out[0], skip_special_tokens=True))
```

### Chat template (multi-turn)

```python
messages = [
    {"role": "user", "content": "What is an impulse in Edge Impulse?"}
]
formatted = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(formatted, return_tensors="pt").to(model.device)

with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=512,
                         do_sample=True, temperature=0.7,
                         pad_token_id=tokenizer.eos_token_id)

print(tokenizer.decode(out[0], skip_special_tokens=True))
```

---

## Example Prompts

| Task | Prompt |
|---|---|
| Concept explanation | `"What is a DSP block in Edge Impulse?"` |
| API usage | `"How do I use the Edge Impulse Python SDK to upload data?"` |
| Deployment | `"How do I deploy a model to an Arduino Nano 33 BLE Sense?"` |
| Code generation | `"Write Python code to collect IMU data and upload it to Edge Impulse."` |
| Troubleshooting | `"Why is my Edge Impulse model showing high latency on the Cortex-M4?"` |

---

## Limitations

- Based on a 0.5B parameter base model — may struggle with complex multi-step reasoning
- Training data covers documentation as of mid-2026; newer APIs may not be represented
- May hallucinate undocumented Edge Impulse features
- Not suitable for safety-critical or production decision-making systems
- Validate generated code before running on real hardware

---

## Related Models

| Model | Base | Status |
|---|---|---|
| This model | Qwen1.5-0.5B | ✅ Available |
| [eoinedge/edgeai-qwen2.5coder-1.5b-lora](https://huggingface.co/eoinedge/edgeai-qwen2.5coder-1.5b-lora) | Qwen2.5-Coder-1.5B-Instruct | 🔄 Training |
| [eoinedge/edgeai-qwen2.5coder-7b-lora](https://huggingface.co/eoinedge/edgeai-qwen2.5coder-7b-lora) | Qwen2.5-Coder-7B-Instruct | 🔄 Training |
| [eoinedge/arduino-qwen0.5-lora](https://huggingface.co/eoinedge/arduino-qwen0.5-lora) | Qwen1.5-0.5B | ✅ Available (Arduino docs) |

---

## Citation

```bibtex
@misc{edgeai-docs-embedding-qwen1.5-0.5b-instruct,
  author       = {Jordan, Eoin},
  title        = {edgeai-docs-embedding-qwen1.5-0.5b-instruct},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/eoinedge/edgeai-docs-embedding-qwen1.5-0.5b-instruct}}
}
```