Modelpublic

LL-Square/CodeForge-TinyLlama1.1B-Instruct

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

0likes4downloads

Model Card

CodeForge-Instruct

Lightweight repository for preparing, training, and uploading small instruct-style models and LoRA adapters.

This project contains simple scripts to train a model (train.py), run inference (main.py), configure logging (logging_setup.py), and upload artifacts (upload.py). A small sample dataset is included as sample.jsonl.

Data format

The dataset expects newline-delimited JSON (.jsonl) where each line is an object with at least prompt and response (or instruction/output) fields. Example (sample.jsonl):

jsonl

{"prompt": "Summarize the following text:", "response": "A short summary."}

Adjust train.py to match your field names if needed.

Usage

Training (example):

bash

python train.py --data sample.jsonl --output-dir ./checkpoints --epochs 3 --batch-size 8

Run inference/demo:

bash

python main.py --model ./checkpoints/latest

Upload artifacts (example):

bash

python upload.py --model ./checkpoints/latest --dest hub-or-bucket

See individual scripts for additional flags and configuration.

Logging

The repository centralizes logging in logging_setup.py; import and call setup_logging() from other scripts to get consistent formatting and levels.

Development

—Run linters/formatters as you prefer (e.g. black, ruff).
—Add tests under a tests/ folder if you expand behavior.

Contributing

Open issues or PRs with clear reproduction steps. Keep changes minimal and scoped.

License

This repository does not include a license file. Add a LICENSE if you plan to publish.

Framework versions

—PEFT 0.18.1
— --- If you'd like, I can:
—add a requirements.txt with pinned versions,
—add CLI argument parsing examples to train.py and main.py, or
—create a short CONTRIBUTING guide.