weft-lineage-extractor-3b

weft-lineage-extractor-3b

The efficient tier of the Weft lineage-extractor family: a 3B code model (LoRA fine-tuned, merged) that extracts table- and column-level data lineage from ETL scripts as structured JSON. Deployable on a single 12 GB consumer GPU. Trained on real-world GitHub ETL scripts with tri-vendor consensus silver labels.

This repository ships three branches covering the table↔column trade-off frontier:

Benchmark: 129 non-empty real GitHub scripts, tri-vendor consensus gold, greedy decoding at 1024 max new tokens. Column metrics are conditional on matched tables.

Which branch? One model for both tables and columns → main. Maximum column quality (pair it with a table specialist via inference-time fusion) → tri-column-specialist. Maximum table recall → tri-table-specialist.

Method highlight: table-token loss weighting

The 3B table↔column frontier is driven by gradient imbalance, not just capacity: in the answer JSON, high-entropy column tokens outnumber table tokens 4.39 : 1, drowning the table-name gradient. Re-weighting the loss on table-structure tokens (W=3) lifts table recall 0.645 → 0.734 with no data removal and no extra capacity, at table-precision parity (0.834, McNemar n.s.). Loss weighting strictly dominates column-density dilution: at equal table recall it keeps ~+0.15 column F1 that dilution would destroy.

Label credibility (tri-vendor consensus)

Gold and silver labels are 2-of-3 consensus across three independent vendors (qwen-max ∩ deepseek-v4-pro ∩ GPT-5.6). GPT-5.6, which never participated in constructing the earlier two-vendor labels, independently agrees with them at 0.976 (table) / 0.958 (column); models trained only on the two-vendor subset reach 0.782 table recall on edges only GPT-5.6 confirms — evidence the model learns real lineage, not one vendor's labeling habits. Total teacher-labeling cost: $2.42.

Honest boundaries

• Reduced, not eliminated, circularity: labels remain LLM-consensus silver; no human gold.
• Governance routing: 3-of-3 vendor-unanimous cases (~70%) are candidates for an auto-adopt layer; vendor-disagreement cases (~30%) route to human review. The model narrows the review queue; it does not eliminate review.
• Convention A exclusions (dynamically-built table names, commented-out/printed SQL, temp views) are deliberate scope boundaries of static extraction, not bugs.
• The strict dual gate (table R ≥ 0.75 and column F1 ≥ 0.85) is unreachable at 3B — shown twice independently (loss-weight sweep; r=64 capacity stack). It remains unreachable by any single model even at 14B; the quality path is weft-lineage-extractor-14b (best balanced 0.799 / 0.856) plus inference-time dual-expert fusion.

Training details

Usage

vLLM (recommended for batch extraction — continuous batching, OpenAI-compatible)

pip install vllm
vllm serve wallfacers/weft-lineage-extractor-3b --revision main \
  --dtype bfloat16 --max-model-len 2048 --gpu-memory-utilization 0.9 --port 8000

from vllm import LLM, SamplingParams
llm = LLM(model="wallfacers/weft-lineage-extractor-3b", revision="main",
          dtype="bfloat16", max_model_len=2048)
sp = SamplingParams(temperature=0.0, max_tokens=256)      # deterministic decoding
outs = llm.chat([[{"role": "system", "content": SYSTEM_PROMPT},
                  {"role": "user", "content": "task_type: PYTHON\nscript:\n" + script}]], sp)
print(outs[0].outputs[0].text)

transformers (single request / interactive, p50 ≈ 360 ms)

from transformers import AutoModelForCausalLM, AutoTokenizer
REPO, REV = "wallfacers/weft-lineage-extractor-3b", "main"
tok = AutoTokenizer.from_pretrained(REPO, revision=REV)
model = AutoModelForCausalLM.from_pretrained(REPO, revision=REV,
                                             dtype="bfloat16", device_map="cuda")

System prompt (shared across the family):

You are a data lineage extractor for ETL scripts. Given a PYTHON, SHELL, SCALA or JAVA task
script (Spark/Flink jobs included), output ONLY a JSON object {"reads": [...], "writes": [...]}
where each item is {"table": str, "columns": [str] or null}. Rules: include a table only if
its literal name appears in the script text; ignore dynamically-built table names,
commented-out SQL, and SQL that is merely printed or logged; if nothing is read or written,
output {"reads": [], "writes": []}.

Output schema: {"reads": [{"table": str, "columns": [str] | null}], "writes": [...]}.

Cost note: lineage answers are short (~22 output tokens/request, prefill-bound). Self-hosted on a single consumer GPU the marginal cost approaches electricity (≈$0.004 / 1k requests) — 1–2 orders of magnitude below cloud LLM APIs for high-volume batch extraction. Full throughput/cost ledger: out/cost-analysis-068.md in the GitHub repo.

Model family

Full evidence ledger (4.39:1 token measurement, frontier-dominance +0.15, McNemar parity, dual-gate negative results): github.com/wallfacers/data-weave (ml/lineage-extractor/out/PAPER-EVIDENCE-068.md).