Up and running in Hugging Face Space using 2 virtual cpu's and 16 GB RAM!! CFR-FineTuned_III

Llama-3.2-1B Fine-tuned on the Code of Federal Regulations (CFR)

This is a fine-tuned version of meta-llama/Meta-Llama-3.2-1B trained on all sections from the United States Code of Federal Regulations (CFR). The goal: provide a specialized assistant for navigating and answering questions about U.S. federal regulations.

Model Description

Base Model: Llama-3.2-1B
Method: QLoRA, 4-bit quantization
Dataset: Custom, parsed from CFR XML (Titles 1-50)
Epochs: 3
Tokens Seen: ~243M
Final Training Loss: 1.267
Mean Token Accuracy: 0.739
Training Time: ~5h 17m

Hardware/Environment:
Training was conducted on Modal using a single NVIDIA H200 GPU.
Training speed: ~1.10 steps/sec, 35 samples/sec.

Note: This loss is typical for a Llama-3 1B model on legal/complex text. For comparison: random output would yield >2.0; perfect memorization of a small dataset would yield <1.0. This is in the “actually learned something useful” range for this setup.

Intended Uses & Limitations

Intended Uses

Regulatory Q&A
Summarization of CFR text
Text generation related to U.S. federal regulations

Limitations

NOT a substitute for legal advice. Output may be incorrect or outdated (data as of 2024-06-25).
Can hallucinate—don’t trust answers without checking against the source.
Validation/test loss is not reported here (evaluate on your own task/data before using in production).

How to Use

You can use this model directly with the transformers library.