ihounie/huggy-2-1k-alpaca

ihounie/huggy-2-1k-alpaca

This repository contains LoRA adapters trained on an Alpaca-style dataset (1k longest unsafe responses filtering used in training config).

• Base model: huggyllama/llama-7b
• Artifact source (Weights & Biases): alelab/SAFE-llama2-long1k/30xn8tf4-lora_adapters:latest
• Exported at: 2026-02-05T20:45:10Z
• Experiment: safety
• Global step: 80
• Training output_dir: ./outputs/safe/erm/baseline

Usage (load adapters)

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained(
    "huggyllama/llama-7b",
    device_map="auto",
)
tok = AutoTokenizer.from_pretrained("huggyllama/llama-7b", use_fast=True)

model = PeftModel.from_pretrained(base_model, "ihounie/huggy-2-1k-alpaca")

Merge adapters into a standalone model

This repo ships a helper script:

python merge_and_save.py --base_model "huggyllama/llama-7b" --adapter_dir . --out_dir ./merged

The merged folder can then be uploaded as a fully standalone model (no PEFT dependency).

Files

• adapter_config.json, adapter_model.*: LoRA adapter weights/config
• Tokenizer files (if present): tokenizer.*, special_tokens_map.json, etc.