BeastxD/text2cypher_lora_v8_raw

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 22, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The BeastxD/text2cypher_lora_v8_raw model is a 4.5 billion parameter Qwen3.5-4B fine-tuned (LoRA, merged to 16-bit) model developed by BeastxD. It specializes in translating natural language questions and a graph schema into Cypher queries for Neo4j graph databases. This model is optimized for generating accurate Cypher queries, achieving 53.18% execution ExactMatch on the neo4j/text2cypher-2024v1 test set. Its primary use case is enabling natural language interaction with graph databases by converting user queries into executable Cypher.

Loading preview...

Model Overview

BeastxD/text2cypher_lora_v8_raw is a 4.5 billion parameter model based on Qwen/Qwen3.5-4B, fine-tuned using LoRA (merged to 16-bit) to convert natural language questions and a provided graph schema into Cypher queries for Neo4j graph databases. This model is designed to work with arbitrary, unseen schemas, which must be supplied in the system prompt.

Key Features and Usage

  • Cypher Query Generation: Translates natural language questions into Cypher queries, utilizing a provided graph schema as the single source of truth.
  • Schema-Dependent: Requires the graph schema to be passed in the system prompt; it is not baked into the model weights.
  • Optimized for Qwen3.5: Specifically fine-tuned on Qwen3.5-4B with custom LoRA target modules to account for its Gated DeltaNet architecture, which differs from default Unsloth configurations.
  • Training Data: Trained on neo4j/text2cypher-2025v1 with 34,323 training rows and 869 schemas, using raw labels.
  • Performance: Achieved an execution ExactMatch of 53.18% on the neo4j/text2cypher-2024v1 test set, outperforming a neo4j/text2cypher-gemma class third-party baseline (42.25%) and GPT-4o (~30%) on the same harness.

Important Usage Notes

  • Disable Thinking: Users must set enable_thinking=False when generating output, as the model was trained with empty <think></think> blocks.
  • Prune Schema: It is crucial to prune the schema to include only elements relevant to the question, as the model was trained on reduced schemas.

Limitations

  • Raw Labels: Trained on raw labels, which include some mutually contradictory gold Cypher queries in the training data.
  • Understated Score: The execution ExactMatch score may be pessimistic due to projection mismatches (e.g., gold returns a node, model returns properties).
  • Property Selection: Known issue with selecting the wrong property when multiple are plausible.
  • English Only: Trained and evaluated exclusively on English questions.

Finetunes
1 models