wallfacers/weft-lineage-extractor-3b

TEXT GENERATIONConcurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 5, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

The wallfacers/weft-lineage-extractor-3b is a 3 billion parameter model, fine-tuned from Qwen/Qwen2.5-Coder-3B-Instruct, designed for extracting table-level data lineage (reads/writes) from ETL scripts as JSON. It was fine-tuned using LoRA on real open-source ETL scripts, achieving a table precision of 0.64 on a held-out benchmark. This model is optimized for self-hostable, cost-free lineage extraction for imperative/SQL-embedding ETL scripts, matching frontier LLM performance at a smaller scale.

Loading preview...

wallfacers/weft-lineage-extractor-3b: Precision Lineage Extraction

This 3 billion parameter model, fine-tuned from Qwen/Qwen2.5-Coder-3B-Instruct, specializes in extracting table-level data lineage (reads/writes) from ETL scripts. Unlike its synthetic-trained predecessors, this model was fine-tuned using LoRA on approximately 1,774 real silver-labelled ETL scripts (Python/Shell/SQL) from bigcode/the-stack-dedup.

Key Capabilities & Differentiators

  • High Precision Lineage: Achieves a table precision of 0.64 (with calibrated gold and grounding filter) on a held-out, source-isolated benchmark, significantly improving upon previous synthetic-trained models (0.33 precision).
  • Self-Hostable: Designed to run efficiently on a single 12 GB GPU, making it suitable for self-hosted, cost-free deployment.
  • JSON Output: Outputs lineage in a structured JSON format {"reads": [...], "writes": [...]}.
  • Precision-First Design: Tuned for precision, which is critical for data governance, even at the cost of some recall. It deliberately excludes dynamically-built table names, commented SQL, and temporary views.
  • Grounding Filter: Includes a deterministic grounding filter in its recommended inference recipe, which enforces that a table is counted only if its literal name appears in the script, boosting precision.

Intended Use Cases

  • Cost-free, self-hosted table-level lineage for imperative/SQL-embedding ETL scripts.
  • As the LLM channel within a routed lineage system, complementing rule parsers and SQL parsers.
  • For scenarios requiring a precision-first extractor, especially when paired with the grounding filter to ensure literal table name presence.

Limitations

  • Does not provide column-level lineage.
  • Excludes dynamically-built table names and commented/logged SQL.
  • Not intended for ground-truth-critical governance without human review due to inherent ambiguities in some real scripts.

Finetunes
1 models