TinyDevAssist LoRA adapter

This is a PEFT LoRA adapter for HuggingFaceTB/SmolLM2-360M-Instruct at upstream revision a10cc1512eabd3dde888204e902eca88bddb4951. It is an educational developer-support experiment, not a production assistant.

The project-authored Apache-2.0 seed contains 36 English question-answer pairs across Git, Python, HTTP, APIs, debugging, and operations. A deterministic seed-42 split provided 24 train, 6 validation, and 6 test records. Training used one epoch and six optimizer steps with rank 8, alpha 16, dropout 0.05, and q_proj/v_proj targets. The CPU run used Python 3.11.16, PyTorch 2.14.1+cpu, Transformers 4.57.6, and PEFT 0.21.2. Training loss was 3.27303; validation loss was 3.39398.

A paired generation comparison on the six held-out prompts produced nonempty responses from both base and adapter. Mean generation latency on that CPU run was 9.520 seconds for the base and 10.301 seconds for the adapter. No independent quality score or improvement claim is available. The small authored dataset cannot establish generalization. Responses may contain wrong or unsafe commands; verify them before use. Text inspection of the six held-out responses found a fabricated psutil API, misleading Git guidance, and premature operations advice. See the review for the exact cases. These observations do not establish a population error rate.

Reproduce and use

Try the public Gradio ZeroGPU demo. One anonymous hosted inference smoke test returned a nonempty response; this does not establish sustained availability or answer quality. Free ZeroGPU requests may be queued or limited by quota.

Clone the source repository, install Python 3.11 dependencies with python -m pip install -e '.[app,dev]', and set HF_MODEL_REPO to the exact published adapter repository ID before running python app.py. The demo loads the base at the pinned revision above, applies this PEFT adapter, and serves Gradio on port 7860. Its tokenized prompt plus requested new tokens must fit the 512-token training sequence budget.

For the reported experiment, regenerate the seed-42 split with python prepare_data.py --seed 42. The source SHA-256 is 92ffc379ffe09d71052d9a7cfc9c9b4fe740c540cb7955e4347e2f646bef151f; the held-out split SHA-256 is 706756348bd549ee3db4f453189955b84950760a723751fd287b37134566dde7. The evaluation protocol explains how to rerun both model paths and inspect the six response pairs.

Load this adapter with PEFT on the pinned base revision. See the source repository for the data card, scripts, configuration, tests, and evaluation protocol.