neuralbroker/blitzkode-lora-0.5b

BlitzKode LoRA Adapter (0.5B)

BlitzKode is a local AI coding assistant fine-tuned from Qwen/Qwen2.5-0.5B-Instruct using LoRA (Low-Rank Adaptation). This repository contains the PEFT adapter — the research-friendly version that can be hot-loaded on top of the base model.

Creator: Sajad (neuralbroker) GitHub: https://github.com/neuralbroker/blitzkode Production GGUF: neuralbroker/blitzkode

---

Model Details

Property	Value
Adapter version	2.1
Base model	Qwen/Qwen2.5-0.5B-Instruct
LoRA rank (r)	16
LoRA alpha	32
LoRA dropout	0.05
Target modules	v_proj, up_proj, down_proj, q_proj, o_proj, gate_proj, k_proj
Training steps	50
Final loss	~0.48
Library	PEFT
License	MIT

---

Training Pipeline

This adapter was produced by a 4-stage fine-tuning pipeline applied to the Qwen2.5 family:

Stage	Method	Purpose
1	SFT	Supervised fine-tuning on 71 curated algorithmic coding problems
2	Reward-SFT	Continued SFT with heuristic reward signals for code correctness and formatting
3	DPO	Direct Preference Optimization on handcrafted chosen/rejected pairs
4	LoRA SFT (this adapter)	Final LoRA fine-tune (r=16) on 99 samples; base model Qwen2.5-0.5B

Training Dataset (199 total samples)

Subset	Count	Source	License
Curated algorithmic problems	71	Custom (local) — arrays, strings, trees, DP, graphs	MIT
MetaMathQA samples	100	meta-math/MetaMathQA	CC BY 4.0
Python/JavaScript patterns	28	Custom (local) — decorators, context managers, data classes	MIT
Total	199

---

Usage

Load with PEFT

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model_id = "Qwen/Qwen2.5-0.5B-Instruct"
adapter_repo  = "neuralbroker/blitzkode-lora-0.5b"

tokenizer = AutoTokenizer.from_pretrained(base_model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(model, adapter_repo)
model.eval()

Generate code

prompt = (
    "<|im_start|>system\n"
    "You are BlitzKode, a precise AI coding assistant created by Sajad.\n"
    "<|im_end|>\n"
    "<|im_start|>user\n"
    "Write a Python function for binary search with full edge-case handling.\n"
    "<|im_end|>\n"
    "<|im_start|>assistant\n"
)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=300,
    temperature=0.7,
    do_sample=True,
    repetition_penalty=1.1,
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Merge adapter into base model (for export)

merged = model.merge_and_unload()
merged.save_pretrained("blitzkode-0.5b-merged")
tokenizer.save_pretrained("blitzkode-0.5b-merged")

---

Prompt Format

BlitzKode uses the ChatML template standard for Qwen models:

<|im_start|>system
You are BlitzKode, a precise AI coding assistant created by Sajad.<|im_end|>
<|im_start|>user
{your question}<|im_end|>
<|im_start|>assistant

---

Limitations

• Text-only — no image/multimodal support.
• 0.5B parameters — smaller and faster than the 1.5B GGUF variant; may be less accurate on complex algorithmic tasks.
• 2048-token context — not suitable for long repository-level analysis.
• Review all outputs — generated code must be tested before use in production.
• Not security-audited — do not use for cryptographic or safety-critical code without thorough expert review.
• Math reasoning — MetaMathQA training improves basic reasoning but does not substitute a dedicated math model.

---

Relation to the Production Model

Variant	Repo	Size	Runtime	Use case
GGUF (1.5B, F16)	neuralbroker/blitzkode	~3 GB	llama.cpp / llama-cpp-python	Production; CPU/GPU, no Python ML stack needed
LoRA adapter (0.5B)	neuralbroker/blitzkode-lora-0.5b (this repo)	~100 MB	PEFT + Transformers	Research; merging, further fine-tuning, quantization

---

License

MIT — see LICENSE.

You must also comply with the upstream Qwen/Qwen2.5-0.5B-Instruct license when redistributing any derived weights.

---

Citation

@software{blitzkode2025,
  author  = {Sajad},
  title   = {BlitzKode: A Local AI Coding Assistant},
  year    = {2025},
  url     = {https://github.com/neuralbroker/blitzkode}
}