neuralbroker/blitzkode-1.5b-lora

BlitzKode 1.5B LoRA Adapter

This repository contains the BlitzKode 1.5B PEFT/LoRA adapter for Qwen/Qwen2.5-1.5B-Instruct. It is intended for research, inspection, and reproducible adapter loading. The production local-inference artifact is the merged GGUF model at neuralbroker/blitzkode.

Intended use

• Local coding-assistant research
• Adapter inspection and continued fine-tuning
• Reproducing the BlitzKode GGUF export pipeline
• Educational experiments with PEFT/LoRA on Qwen2.5

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/Qwen2.5-1.5B-Instruct"
adapter_id = "neuralbroker/blitzkode-1.5b-lora"

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

BlitzKode was developed through staged coding-assistant post-training:

Stage	Method	Purpose
SFT	Curated coding examples	Teach concise coding responses and common algorithms
Reward-SFT	Heuristic continuation	Reinforce formatting, correctness patterns, and practical explanations
DPO	Preference pairs	Penalize weak answers and hallucinated details
Export	Merge + GGUF	Produce the production blitzkode.gguf artifact

Dataset provenance is tracked in the GitHub project and the production model repository docs.

Evaluation

The latest 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 is a lightweight regression smoke eval, not a comprehensive benchmark such as HumanEval, MBPP, or SWE-bench.

Limitations

• Small-model limitations apply; outputs can be wrong or incomplete.
• Direct prompting may hallucinate unsupported APIs or signatures.
• Default deployment context is 2,048 tokens.
• Text-only model; no image/file multimodal support.

Related repositories

• Production GGUF/API docs: neuralbroker/blitzkode
• Lightweight 0.5B adapter: neuralbroker/blitzkode-lora-0.5b
• 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.