Aswinkv07

qwen3.5-4b-neo4j-text2cypher-lora

Deploy Dedicated

README

License: apache-2.0

Intended use

The model expects:

A Neo4j graph schema containing the available labels, relationship types, and properties.
A natural-language question about that graph.

It is trained to return one Cypher query without Markdown or explanatory text.

Training details

Base model: unsloth/Qwen3.5-4B
Dataset: neo4j/text2cypher-2025v1
Method: supervised fine-tuning with LoRA using Unsloth, TRL, and PEFT
Training steps: 2,200 (approximately one epoch)
Maximum sequence length: 4,096 tokens
Per-device batch size: 2
Gradient accumulation: 8
Effective batch size: 16 on one GPU
Learning rate: 2e-4 with cosine scheduling
Warmup ratio: 0.03
Weight decay: 0.01
LoRA rank / alpha / dropout: 32 / 64 / 0.0
LoRA targets: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, and down_proj
Training hardware: NVIDIA L40S

Usage with Unsloth

python
from unsloth import FastLanguageModel

adapter_id = "Aswinkv07/qwen3.5-4b-neo4j-text2cypher-lora"

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=adapter_id,
    max_seq_length=4096,
)
FastLanguageModel.for_inference(model)

schema = '''Node properties:
- Person: name (STRING)
- Movie: title (STRING)
Relationships:
- (:Person)-[:ACTED_IN]->(:Movie)'''
question = "Which movies did Tom Hanks act in?"

messages = [
    {
        "role": "system",
        "content": (
            "You are an expert Neo4j Cypher query generator. "
            "Given a graph schema and a natural language question, generate "
            "exactly one valid Cypher query. Return only the Cypher query. "
            "Do not include markdown, explanations, comments, or extra text. "
            "Use only labels, relationship types, and properties that appear "
            "in the provided schema."
        ),
    },
    {
        "role": "user",
        "content": (
            f"Graph schema:\n{schema}\n\n"
            f"Question:\n{question}\n\n"
            "Generate the Cypher query."
        ),
    },
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(
    input_ids=inputs,
    max_new_tokens=512,
    do_sample=False,
)
query = tokenizer.decode(outputs[0, inputs.shape[-1]:], skip_special_tokens=True)
print(query.strip())

Evaluation

No formal benchmark result for the final step-2,200 adapter is reported in this model card yet. Evaluate generated queries using syntax validity, schema adherence, and execution-result accuracy. Exact-string matching alone can mark semantically equivalent Cypher queries as incorrect.

Limitations and safety

The model can produce invalid Cypher or reference the schema incorrectly.
Performance may vary across Neo4j versions, schemas, and query complexity.
The training data is primarily English and may not represent every graph domain.
Validate generated queries before execution. For production use, apply timeouts and least-privilege or read-only database credentials.

License

The adapter is released under Apache-2.0, following the base model license. Users must also comply with the licenses and terms of the base model and training dataset.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

Aswinkv07

Model Tree

Base

unsloth/Qwen3.5-4B

Adapter

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality