LLaDA-8B-Bioinspired-dLLM-Instruct-11-21-2025

LLaDA-8B-Bioinspired-dLLM-Instruct

This model is a fine-tuned version of GSAI-ML/LLaDA-8B-Instruct on a a dataset of bio-inspired materials.

Inference examples

Make sure to install dLLM:

git clone https://github.com/ZHZisZZ/dllm.git
cd dllm
pip install -e .

Simple generation

Example inference:

from dataclasses import dataclass
import transformers
import dllm
from dllm.tools.chat import decode_trim
from dllm.pipelines import llada

''' #or log in using `huggingface-cli login`
token= 'hf_...'  
from huggingface_hub import login
login(token=token)
'''

# ---------------------------------------------------------
# Load model + tokenizer
# ---------------------------------------------------------

@dataclass
class ScriptArguments:
    model_name_or_path: str = "lamm-mit/LLaDA-8B-Bioinspired-dLLM-Instruct-11-21-2025"

    def __post_init__(self):
        self.model_name_or_path = dllm.utils.resolve_with_base_env(
            self.model_name_or_path, "BASE_MODELS_DIR"
        )

script_args = ScriptArguments()

transformers.set_seed(42)

model = dllm.utils.get_model(model_args=script_args).eval()
tokenizer = dllm.utils.get_tokenizer(model_args=script_args)

generator = llada.LLaDAGenerator(
    model=model,
    tokenizer=tokenizer,
)

gen_config = llada.LLaDAGeneratorConfig(
    steps=256,
    max_new_tokens=256,
    block_length=32,
    temperature=0.0,
    remasking="low_confidence",
)

# ---------------------------------------------------------
# Batched inference step
# ---------------------------------------------------------

messages_batch = [
    [{"role": "user", "content": "Explain materiomics briefly."}],
    [{"role": "user", "content": "Define mechanobiology in one paragraph."}],
    [{"role": "user", "content": "Why is silk stronger than elastin?"}],
]

inputs = tokenizer.apply_chat_template(
    messages_batch,
    add_generation_prompt=True,
    tokenize=True,
)

outputs = generator.generate(
    inputs,
    gen_config,
    return_dict_in_generate=True,
)

sequences = decode_trim(tokenizer, outputs.sequences.tolist(), inputs)

# ---------------------------------------------------------
# Results
# ---------------------------------------------------------

for i, s in enumerate(sequences):
    print("\n" + "-" * 70)
    print(f"[Sample {i}]")
    print("-" * 70)
    print(s.strip())

Visualization:

terminal_visualizer = dllm.core.generation.visualizer.TerminalVisualizer(
        tokenizer=tokenizer
    )
terminal_visualizer.visualize(outputs.histories, rich=True)

Infill generation

Example to extract reasoning and design principles:

gen_config = llada.LLaDAGeneratorConfig(
    steps=512,
    max_new_tokens=512,
    block_length=32,
    temperature=0.2,
    remasking="low_confidence",
)
masked_messages = [
    [
        {
            "role": "user",
            "content": (
                "In spider-silk materiomics, we often optimize hierarchical structure "
                "from amino-acid sequence to β-sheet nanocrystal arrangement. "
                "Complete the missing reasoning steps for the following design question:\n\n"
                f"**Design Problem:** How could one tune the fraction of β-sheet "
                f"nanocrystals to increase toughness without compromising elasticity?\n\n"
                f"Missing reasoning: {tokenizer.mask_token * 128}"
            ),
        },
        {
            "role": "assistant",
            "content": (
                f"The summary is: {tokenizer.mask_token * 20}"  #  
            ),
        },
    ],

    [
    {
        "role": "user",
        "content": (
            "In nacre-inspired composite design, we often tune the architecture of "
            "brick-and-mortar layers to balance stiffness, strength, and toughness. "
            "Complete the missing reasoning steps for the following design question:\n\n"
            "**Design Problem:** How could one introduce controlled mineral platelet "
            "misalignment to enhance toughness while preserving high stiffness?\n\n"
            f"Missing reasoning: {tokenizer.mask_token * 128}"
        ),
    },
    {
        "role": "assistant",
        "content": (
            f"The design principle is: {tokenizer.mask_token * 20}"
        ),
    },
]

]

# Tokenize input with NO generation prompt
inputs = tokenizer.apply_chat_template(
    masked_messages,
    add_generation_prompt=False,
    tokenize=True,
)

# Infilling
outputs = generator.infill(inputs, gen_config, return_dict_in_generate=True)
sequences = decode_trim(tokenizer, outputs.sequences.tolist(), inputs)

# Print results
for idx, (inp, filled) in enumerate(zip(inputs, sequences)):
    print("\n" + "-" * 80)
    print(f"[Case {idx}]")
    print("-" * 80)
    print("[Masked]:\n" + tokenizer.decode(inp))
    print("\n[Filled]:\n" + (filled.strip() if filled.strip() else "<empty>"))

print("\n" + "=" * 80 + "\n")

terminal_visualizer.visualize(outputs.histories, rich=True)