Modelpublic

kparvataneni/llama31-8b-instruct-sft-balanced-2k-phi-benign

sourceHugging Facellama3.1updated 1mo agoView on Hugging Face

0likes7downloads

Model Card

Llama 3.1 8B Instruct — balanced 2k PHI–benign SFT

Synthetic twin-dossier SFT used for MemoDiff PHI vs benign memorization experiments (`kveni12/model-memo-diff`).

Checkpoints

Each training step is a revision / branch named step-XXXXXX (zero-padded), same convention as PHIMemo/llama31-8b-instruct-sft-balanced-3k.

Available steps: 30, 60, 120, 180, 240, 300, 360, 480, 540, 570, 600.

python

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "kparvataneni/llama31-8b-instruct-sft-balanced-2k-phi-benign"
tok = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
model = AutoModelForCausalLM.from_pretrained(repo, revision="step-000600")

Tokenizer files are not bundled; use the base Instruct tokenizer.