adapter_model.safetensors: LoRA adapter weights
adapter_config.json: PEFT adapter configuration
tokenizer.json, tokenizer_config.json, special_tokens_map.json: tokenizer files
chat_template.jinja: chat template used with the checkpoint
training_args.bin: serialized training arguments
metadata.json: structured provenance metadata
SOURCE_RUN.txt: source training run provenance

Usage

python

from transformers import AutoModel, AutoTokenizer
from peft import PeftModel

base_model_id = "GSAI-ML/LLaDA-8B-Instruct"
adapter_id = "OnAnOrange/llada-8b-cora-lp-notopo-2hop-r64-ep5"

tokenizer = AutoTokenizer.from_pretrained(adapter_id, trust_remote_code=True)
base_model = AutoModel.from_pretrained(
    base_model_id,
    trust_remote_code=True,
    torch_dtype="auto",
)
model = PeftModel.from_pretrained(base_model, adapter_id)

The exact inference wrapper depends on the DLM-Graph/LLaDA evaluation pipeline. This repository publishes the trained adapter and provenance files, not a standalone merged full model.

Provenance

Original local training run:

tmdlm-llada-8b-cora-lp-2hop-notopo-r64-ep5-cora_lp_llaga_notopo_20260523_0829_8gpu_5ep

The uploaded checkpoint is checkpoint-final from that run. See metadata.json and SOURCE_RUN.txt for the exact local source path.

Limitations

This checkpoint is specialized for DLM-Graph experiments. It is not intended as a general-purpose instruction model. Users should evaluate it within the same graph prompting and decoding setup used by the DLM-Graph experiments.