gears-div9-estimator-gemma4-31b

Generic Gears Gym — Div-9 Estimator (QLoRA adapter)

A vertical-specialized construction estimating model: given a fitness-club project's scoped quantities (flooring takeoff line items with units), it produces priced line items and a bid total consistent with a real commercial flooring subcontractor's cost structure.

This is a research artifact from an applied-AI program studying whether a subcontractor's bid history is a trainable asset. The training corpus is a single US flooring subcontractor's verified fitness-vertical bid archive (39 projects, 2023–2026), with every client, brand, and company identifier replaced before training — the weights never saw a real name. Dollar figures and trade content are intentionally preserved: pricing is the task.

Model details

• Base: google/gemma-4-31b-it, 4-bit NF4 (QLoRA)
• Adapter: r=8, α=160, top-32-of-60 decoder layers, 7 projection targets (32.7M trainable params, 0.10%)
• Training: 6 epochs, lr 1e-5 constant, seq 6144, effective batch 2, full-sequence loss; 3 seeds trained, best-by-holdout promoted
• Data: 39 sanitized fitness-vertical estimate pairs (train) + 4 val + 4 held-out projects for generation eval

Evaluation

n=4 is a pilot-scale holdout; treat these as descriptive. Notable honest finding: this vertical-only model does not beat the full-corpus parent's fitness-slice performance — 39 vertical rows lose to cross-vertical pricing transfer. Vertical specialization should upweight the vertical within the full corpus rather than isolate it.

Context: the parent full-corpus model scores 13.4–17.3% median APE across seeds on a 50-project mixed-vertical holdout; its fitness-vertical slice scores 1.4–5.0% — the motivation for this vertical specialization.

Weights availability

This repository is card-only by policy: the program's zero-tolerance release bar for identifier emission was not met by the adversarial probe transcripts (prompt-echo strings appear in both tuned and base outputs), so the adapter weights are held privately. The card documents the artifact and its evaluation; weights are not distributed.

Input / output format

Input (user turn): Project: <name>\nGC: <GC-Alpha…>\nDate: <iso>\n\nScoped quantities:\n[{"section":…,"name":…,"qty":…,"unit":…}, …] Output: {"line_items": [{…, "unit_cost": …, "ext": …}], "bid_total": <number>}

Gemma-4 inference note: build generation prompts as an exact prefix of the training-time chat rendering. add_generation_prompt=True primes Gemma-4's thought channel, which this adapter was not trained with, and degrades output format fidelity.

Sanitization & leak posture

• All brand/client/GC/company identifiers replaced with generic tokens (Generic Gears Gym, GC-Alpha…GC-Kappa) prior to training; emails/phones scrubbed; automated zero-leak gates on every dataset file.
• Pre-release red-team (differential): 12 adversarial extraction probes run against the tuned model AND the untuned base as control. The tuned model produced no completions containing real identifiers beyond those the untuned base model produces on identical probes (4/12 probes elicit public-world-knowledge or prompt-echo strings from BOTH models equally). Adapter-attributable leakage: none detected. Probe transcripts necessarily contain the real identifiers they test for, so they stay in the program's internal audit manifest rather than this repo.
• One intentionally held-out sealed project exists for a future public human-vs-model benchmark; it appears in no training or eval split.

Intended use & limitations

Research and demonstration of vertical-specialized estimating. NOT a bidding tool: prices reflect one subcontractor's historical cost structure in one region and era; quantities must come from a real takeoff; no warranty of commercial accuracy. n=39 training projects — expect brittleness outside the fitness-vertical scope distribution.

Provenance & program

Part of an applied research program on markup-as-label training data for construction AI (patent pending; papers in preparation). The open measurement engine this program builds on: OpenTakeoff (Apache-2.0) — npx -y opentakeoff-mcp exposes it to AI agents over MCP. Flagship full-corpus model: div9-flooring-estimator-gemma4-31b. Contact: Kentucky AI.