open-llama-3b-openthought-sft-lora

LoRA adapter for assistant-only SFT on OpenThoughts-114k.

Base model

This adapter is trained on top of ping98k/open-llama-3b-openthought-mid-4bit (the full-loss "mid" model).

Training

• Base: ping98k/open-llama-3b-openthought-mid-4bit
• Method: QLoRA (r=32, alpha=32)
• Data: open-thoughts/OpenThoughts-114k (filtered <= 2024 tokens, 10,582 samples)
• Epochs: 3
• Batch size: 128
• LR: 2e-4 (cosine, warmup 24 steps)
• Loss: assistant-only (system/user turns masked)
• Context: 2024 (native, no RoPE scaling)
• modules_to_save: embed_tokens, lm_head

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

model = AutoModelForCausalLM.from_pretrained("ping98k/open-llama-3b-openthought-mid-4bit", device_map="auto")
model = PeftModel.from_pretrained(model, "ping98k/open-llama-3b-openthought-sft-lora")

Merged model

See ping98k/open-llama-3b-openthought-sft-4bit