scrubdata-qwen3-4b-v6-q8

ScrubData planner v6 — Q8_0 GGUF (the blessed release)

The v6 fine-tune of the ScrubData tabular-cleaning planner (≤4B, plan-emitting, never cell-editing). v6 = the winner of a measured data-scaling loop: same recipe as v5 but with ~3× the real, execution-verified paired supervision (Raha hospital/beers/movies cell-aligned pairs + ROR org-alias vocabulary in the synthetic generator).

Measured (vs v5):

• Real hospital typo repair: recall 0.42 → 0.475, precision 0.16 → 0.185 (frontier zero-shot reference: 0.51)
• Synthetic canonicalization (same protocol): +0.09 F1, seed-robust (3 seeds, non-overlapping CIs)
• Negative results from the same loop (documented in the paper): frequency-derived and algorithm-cleaned labels both reduce quality — real benchmark-gold pairs are what transfer.

Run it (Ollama / llama.cpp):

ollama pull hf.co/ricalanis/scrubdata-qwen3-4b-v6-q8:Q8_0
ollama create scrubdata-ft -f notebooks/Modelfile   # non-thinking template (required)
SCRUBDATA_MODEL=scrubdata-ft uv run server.py

Q8_0 only — Q4_K_M exports corrupt this model (documented). In production the planner is wrapped with reference grounding + calibrated abstention (it never free-generates a canonical for a grounded column type).

Shipped composition (WS1 — verified union planner): in the product, every model-proposed mapping is scored by a deterministic verifier (errors-are-rare frequency gates, variant similarity, reference agreement; threshold SCRUBDATA_TAU, default 0.5) and unioned with the grounded heuristic. Measured on hospital's 509 real errors: 0.905 precision @ 0.413 coverage (gated model plan alone: 0.993 @ 0.287 — 146/147 committed changes correct; seed-robust: 0.891 ± 0.012 @ 0.396 ± 0.025 over 3 training seeds). Dropped merges become review flags, never silent skips.

Post-freeze system (v2, June 2026): the pipeline around this model added four deterministic capabilities (suspect surfacing, generic entity reference, cross-row majority voting, convention gates). On the WildClean benchmark: unseen-source macro F1 0.363 @ damage 0.0219, 0 silent edits across 35 wild tables + 239-table trust audit. Five further fine-tunes and a 3-arm GRPO pilot did NOT improve held-out generalization — v6 remains the champion adapter; the deterministic machinery + verifier carry generalization (central finding, see paper). In the same verify+union harness, two of three zero-shot 24–31B open-weights planners exceed this operating point (0.915 @ 0.485 vs 0.905 @ 0.413, paper §scaling); the 4B remains the most precise gated planner at usable coverage and the only locally-measured one — the architecture, not the fine-tune, is the portable contribution.

Live demo: https://huggingface.co/spaces/build-small-hackathon/scrubdata · Code (open source): https://github.com/ricalanis/scrubdata-hackathon · Paper: docs/paper/ in the repo · Benchmark: https://huggingface.co/datasets/ricalanis/wildclean

Integrity

• 2026-06-12 — scrubdata-qwen3-4b-v6.Q8_0.gguf replaced (same filename, pull commands unchanged); sha256 ef08cc6cc0cfdf289b14459052ab426b66d89c6a0a3c0206bfbdc9f65417e6a1. The previous GGUF (sha256 9caa0b2c52d04bd03dddbddc8452d52a48438d0208cc75371be7728067186ae6) was corrupted by an export bug (declared a Llama-range EOG token id 128247 inside the Qwen3 vocab, degenerating into <tool_call> loops on all runtimes) and has been replaced with a clean re-export from the v6 adapter (peft merge + llama.cpp convert_hf_to_gguf.py --outtype q8_0), smoke-tested before upload. Discovered by our own reproducibility QA.