GRIT
This is an inference checkpoint produced by the GRIT training workflow (Gradient Projection Meets Trust-Region Anchoring for forgetting-resistant reinforcement learning), initialized from Qwen2.5-0.5B-Instruct.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
repository_id = "Duong1211/GRIT"
tokenizer = AutoTokenizer.from_pretrained(repository_id)
model = AutoModelForCausalLM.from_pretrained(
repository_id,
torch_dtype="auto",
device_map="auto",
)
Only inference artifacts are published. Optimizer and trainer-state files are intentionally excluded.