BlitzKode 0.5B LoRA Adapter
This repository contains the lightweight BlitzKode 0.5B PEFT/LoRA adapter for Qwen/Qwen2.5-0.5B-Instruct. It is the smaller research-friendly adapter for low-resource experiments. For the production local API/GGUF release, use neuralbroker/blitzkode.
Intended use
- Lightweight local coding-assistant experimentation
- PEFT/LoRA loading examples
- Continued fine-tuning experiments on small GPUs
- Educational ML engineering workflows
Do not use generated code in production without review and tests.
Loading
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base_model_id = "Qwen/Qwen2.5-0.5B-Instruct"
adapter_id = "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_id)
model.eval()
Training summary
This adapter was trained as a compact LoRA variant for the BlitzKode project. The broader project includes curated coding examples, preference-style data, export scripts, and a production GGUF serving path.
| Property | Value |
|---|---|
| Base model | Qwen/Qwen2.5-0.5B-Instruct |
| Adapter type | PEFT / LoRA |
| Primary task | Coding-assistant text generation |
| Production sibling | neuralbroker/blitzkode GGUF |
Dataset provenance and training scripts are maintained in the GitHub project and production model repository docs.
Evaluation
The current published smoke evaluation is attached to the production GGUF repository: neuralbroker/blitzkode. Current GGUF smoke eval: 3 / 4 passed (75%) on Python factorial, binary search, SQL top users, and fictional-API uncertainty checks. The raw model still fails the fictional-API uncertainty case, so downstream serving should keep guardrails enabled.
This adapter has not been separately benchmarked in this cleanup pass. Treat it as a lightweight research adapter and evaluate it for your target workload before use.
Limitations
- Smaller 0.5B base model has limited reasoning and code-repair capacity.
- Outputs may be syntactically or semantically wrong.
- Direct prompting may hallucinate unsupported APIs or signatures.
- Text-only model; no image/file multimodal support.
Related repositories
- Production GGUF/API docs:
neuralbroker/blitzkode - 1.5B adapter:
neuralbroker/blitzkode-1.5b-lora - GitHub source: https://github.com/neuralbroker/blitzkode
License
MIT for BlitzKode project files and adapter release metadata. You must also comply with the upstream Qwen2.5 license for the base model.
- Downloads last month
- 36