tahamajs / llada-text-to-sql-lora

LLaDA-8B Text-to-SQL (Diffusion-based)

Model Summary

This model is a Text-to-SQL adapter fine-tuned on the GSAI-ML/LLaDA-8B-Instruct base model. Unlike traditional Autoregressive (AR) models that generate tokens left-to-right, this model uses Masked Iterative Generation (Diffusion) .

It treats text generation as a diffusion process: starting with a fully masked sequence and iteratively refining/unmasking tokens based on confidence scores. This allows for bi-directional context utilization during generation.

• Task: Text-to-SQL (Converting natural language questions + schema into SQL queries).
• Method: LLaDA (Large Language Diffusion with Autoregression) with Block Diffusion Sampling.
• Fine-Tuning: QLoRA (4-bit Quantization + LoRA).

Model Details

• Developed by: [Tahamajs/Organization]
• Base Model: GSAI-ML/LLaDA-8B-Instruct
• Dataset: gretelai/synthetic_text_to_sql (Subset of 20k samples)
• Language: English (Natural Language) -> SQL
• Generation Strategy: Semi-Autoregressive / Block Diffusion

How to Use (Inference Code)

Note: This model does not work with the standard model.generate() function because it requires a custom diffusion sampling loop. Use the code below to generate SQL queries.

1. Setup & Loading

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel, PeftConfig

# Device setup
device = "cuda" if torch.cuda.is_available() else "cpu"

# 1. Load Base Model (4-bit)
base_model_id = "GSAI-ML/LLaDA-8B-Instruct"
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.float16,
    bnb_4bit_quant_type="nf4",
)

model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True,
    use_cache=False
)
tokenizer = AutoTokenizer.from_pretrained(base_model_id, trust_remote_code=True)

# 2. Load LoRA Adapter (This Repo)
adapter_model_id = "YOUR_USERNAME/llada-text-to-sql-lora" # Replace with your repo name
model = PeftModel.from_pretrained(model, adapter_model_id)
model.eval()

2. Define Block Diffusion Generation

@torch.no_grad()
def generate_block_diffusion(model, tokenizer, prompt_text, steps=32, gen_len=64):
    """
    Generates text using LLaDA's block diffusion strategy.
    """
    # Tokenize Prompt
    prompt_ids = tokenizer.encode(prompt_text, return_tensors='pt').to(model.device)
    prompt_len = prompt_ids.shape[1]
    
    # Initialize Response with [MASK] tokens
    mask_ids = torch.full((1, gen_len), tokenizer.mask_token_id, device=model.device)
    input_ids = torch.cat([prompt_ids, mask_ids], dim=1)
    
    # Track unknown indices (initially all response tokens)
    unknown_indices = set(range(prompt_len, input_ids.shape[1]))
    tokens_to_lock_per_step = gen_len // steps

    for step in range(steps):
        # Forward pass
        outputs = model(input_ids)
        probs = torch.softmax(outputs.logits, dim=-1)
        
        # Get most confident predictions
        confidences, predicted_ids = torch.max(probs, dim=-1)
        
        # Identify which tokens to "lock in" this step
        candidates = []
        current_unknowns = list(unknown_indices)
        if not current_unknowns: break
        
        for idx in current_unknowns:
            score = confidences[0, idx].item()
            token = predicted_ids[0, idx].item()
            candidates.append((score, idx, token))
            
        # Sort by confidence and pick top k
        candidates.sort(key=lambda x: x[0], reverse=True)
        top_k = candidates[:tokens_to_lock_per_step]
        
        # Update input_ids
        for _, idx, token in top_k:
            input_ids[0, idx] = token
            unknown_indices.remove(idx)
            
    # Decode only the generated part
    return tokenizer.decode(input_ids[0, prompt_len:], skip_special_tokens=True)

3. Run Inference

schema = "CREATE TABLE users (id INTEGER, name TEXT, age INTEGER);"
question = "Show me the names of users older than 25."

prompt = f"""
<|im_start|>system
You are a Text-to-SQL assistant. Output ONLY the SQL query. Do not add explanations.<|im_end|>
<|im_start|>user
Schema:
{schema}

Question:
{question}<|im_end|>
<|im_start|>assistant
"""

output = generate_block_diffusion(model, tokenizer, prompt, steps=32, gen_len=64)
print("Generated SQL:", output)

Training Details

Training Configuration

• Epochs: 5
• Batch Size: 2 (Effective Batch Size = 8 via Gradient Accumulation)
• Optimizer: AdamW (lr=2e-4)
• Scheduler: Linear with Warmup (50 steps)
• Context Length: 384 tokens
• Precision: fp16 (via Mixed Precision)

Noise Schedule

Training used a Forward Masking process where tokens in the answer were randomly replaced with [MASK] based on a uniform time step . Loss was calculated only on masked tokens and reweighted by .

LoRA Configuration

• Rank (r): 16
• Alpha: 32
• Target Modules: q_proj , v_proj
• Dropout: 0.05

Evaluation Results

Evaluated on the gretelai/synthetic_text_to_sql test set (200 samples) using Block Diffusion sampling.

* Scores may vary depending on post-processing strictness and SQL normalization logic.

Bias, Risks, and Limitations

• "Chatty" Output: The model sometimes fails to produce an EOS token immediately after the semicolon, occasionally repeating the query or adding conversational filler. Post-processing (regex extraction of SELECT ... ; ) is recommended.
• Hallucination: In complex queries, the model may occasionally hallucinate columns that do not exist in the provided schema if the schema context is too long or complex.
• Inference Speed: Due to the iterative nature of Block Diffusion (multiple forward passes per generation), inference is slower than standard Autoregressive models of the same size.

Citation

If you use this model or the LLaDA technique, please cite the original paper:

@article{nie2024llada,
  title={LLaDA: Large Language Diffusion with Autoregression},
  author={Nie, Shen and others},
  journal={arXiv preprint arXiv:2402.XXXXX},
  year={2024}
}