Can Small Language Models Run
a Construction Project?

NVIDIA's case is clear: the future of scalable agentic AI is specialised small language models, not ever-larger general-purpose ones. SLMs are 10–30x cheaper to run, can be fine-tuned in hours rather than weeks, run at the edge without cloud dependency, and — when properly trained — produce more reliable structured outputs than frontier models for narrow, repetitive tasks.

The construction industry is a compelling testbed for this thesis. Contract review, delay analysis, and schedule management are high-value, repetitive, and structurally well-defined — exactly the kind of task NVIDIA argues SLMs can own. A specialist construction AI agent running on-device could analyse contracts in real time, flag delay risks before they escalate, and validate schedules without sharing sensitive project data with a cloud provider.

This research programme tests that hypothesis directly: can small language models (3B–20B parameters), fine-tuned on synthetic construction data, reliably perform these tasks? And what are the real limits of the approach?

We evaluated 8 small language models ranging from 3B to 20B parameters on a synthetic construction dataset of 337 examples spanning three domains: contract clause analysis, delay attribution, and schedule analysis. Models span multiple vendors (Mistral, Microsoft, NVIDIA, Google, IBM, Alibaba, OpenAI) and architectures (pure Transformer, hybrid Mamba-2 + Attention, Mixture of Experts). All models were evaluated via the same 6-test standardised evaluation suite (3 Standard + 3 Curveball) with local inference through LM Studio.

Results reveal significant performance variation across the 3B–20B range, with model size showing surprisingly low correlation with performance. The next phase applies a specialist fine-tuning strategy — training each model on a single domain (Contracts, Delays, or Schedules) rather than all three simultaneously — informed by 10 prior trial rounds on Qwen3-4B and Nemotron-3-Nano-4B that demonstrated multi-domain fine-tuning causes catastrophic interference at this dataset scale.

The complete cycle of 8 models × 3 domains × 4 pathways (Base, Base+PE, FT, FT+PE) exposes two architecturally distinct routes to production-grade performance, and the choice between them depends on the knowledge structure of the target domain rather than on model size or training budget.

The first route — fine-tuning — is justified when the target domain admits a closed-form reasoning pattern (TIA framework and Adyard concurrency in Delays; planner-pattern selection and CPM execution in Schedules), when training data of at least sixty pattern-bearing examples is available, and when compute is sufficient for a hundred-plus optimisation steps. Fine-tuned wins in this cycle include Gemma 4 E4B (Contracts +10, Delays +21), Ministral 14B FT+PE (Delays +19), and GPT-OSS 20B FT+PE on Schedules at 92.5/100, the overall best score recorded in the programme.

The second route — a strong base model paired with deterministic scaffolding and an engineered system prompt — is equally viable when resources or training data are constrained, and in several cases is the preferred path even when fine-tuning is feasible. Qwen 3.5 9B Base+PE matched or surpassed its own fine-tune in every domain tested, including a perfect 98/100 on the CB3 Northbrook Solar EPC CPM analysis. The base reasoner, when given explicit rule injection in the system prompt and a deterministic execution layer (CPM solver, clause-database retrieval, validator loop), produces production-grade output without any weight modification, iterates faster, costs less to operate, and adapts dynamically to new clauses, projects, and regulations.

The practical recommendation is therefore domain-conditional rather than model-conditional. Fine-tuning is the preferred path for domains with stable, closed-form knowledge (delay attribution rules, CPM arithmetic, schedule grammar). Base+scaffolding+PE is the preferred path for domains with evolving knowledge (contract clauses, jurisdiction-specific precedent, company-specific commercial terms). The strongest production stack combines both: a fine-tuned planner-analyst model wrapped in retrieval-augmented scaffolding and addressed through a carefully engineered prompt.

We use Unsloth with QLoRA (Quantised LoRA) — a highly optimised fine-tuning framework that reduces VRAM usage and increases training speed 2–5× versus standard PEFT. Base model weights are kept frozen in their original precision; only low-rank adapter matrices (LoRA) are trained. This enables full fine-tuning-quality adaptation with a fraction of the VRAM and time.

• Data format: Chat-template JSONL (system + user + assistant messages)
• Optimiser: AdamW, weight_decay=0.01, cosine LR schedule with warmup
• Gradient accumulation: 4 steps (effective batch size = 4)
• Gradient clipping: max_norm=1.0 (mandatory for SSM stability)
• Evaluation: Per-epoch eval loss on 29-example holdout set
• Checkpointing: Best model by eval loss, with early stopping
• Export: Merged weights → GGUF Q8_0 → LM Studio local API

6 evaluation scenarios across 3 domains — one Standard (T1) and one Curveball (CB) per domain. The T1 tier tests competency on realistic project scenarios; the CB tier tests generalisation to novel jurisdictions, project types, and contract forms entirely unseen in training. Responses are scored by an LLM judge against pre-computed golden answers with full reasoning. All models evaluated across up to 8 pathways (Base/FT × Original/Enhanced prompt × Think/NoThink).

The research programme is structured around four hypotheses, derived from the NVIDIA SLM thesis, ten prior multi-domain trial rounds on Qwen3-4B and Nemotron-3-Nano-4B, and the specialist-strategy revision that followed.

Hypothesis A. Specialist fine-tuning — training each model on a single domain rather than all three simultaneously — will outperform multi-domain fine-tuning by avoiding the catastrophic interference observed in the prior trial rounds. A single-domain adapter learns one task grammar without competing objectives, and the dataset scale used here is insufficient to support concurrent domain mastery.

Hypothesis B. Models that already score highest in a domain's base evaluation possess the strongest foundation for fine-tuning on that domain, and the marginal lift from training should be largest for these top-ranked base models. The expectation is that fine-tuning pushes the leading bases toward the 95–100 range on their specialist domain.

Hypothesis C. Schedules will remain the hardest domain after fine-tuning. Schedule tasks demand multi-step numerical reasoning (CPM forward and backward passes, predecessor arithmetic, lag-type interpretation), and even the strongest base score is only 85.8. Fine-tuning is expected to improve format compliance and pattern recall but is unlikely, on its own, to repair arithmetic reasoning gaps.

Hypothesis D. Model size is not the primary predictor of performance. The base evaluation already shows 4B models outperforming 14B and 20B models on several domains. Architecture (thinking capability, hybrid state-space + attention, mixture-of-experts), pre-training data quality, and fine-tuning data design are expected to matter more than raw parameter count for construction-domain reasoning.

These 8 models represent the latest thinking-capable SLMs in the 3B–14B "Small Language Model" range, with GPT-OSS-20B as the only "small-to-medium" exception. Selection criteria: open-weights, local inference capability on 16 GB VRAM, instruction-following, and structured JSON output support.

The following parameters apply generically across all models during fine-tuning. Specific per-model values will be determined based on base evaluation results and domain assignment.

Phi 4 Mini Reasoning

Strong FIDIC delay causation reasoning (CB2 = 78/100, key discriminator passed). Perfect activity completeness in T3 (25/25) and all durations within benchmark ranges (22/30). The model fails on DB clause ID lookup (all null across both contracts tests), predecessor circular dependencies in T3 (three deadlock chains penalise the C section to 0/25), CPM arithmetic in CB3 (negative total floats, duration overstated), and the FIDIC 19.4 cost rule (Force Majeure is time-only, not cost). Extended thinking (45K–56K tokens) helps classification but does not repair arithmetic or graph construction. Full sub-scores are recorded on the Phi 4 Mini tab.

Nemotron 3 Nano 4B

Contracts score 76/100 combined on clean inputs (T1 = 85, CB1 = 66); the earlier 79 (with "C-002 Rejected, 20/21 correct") came from an answer-leaked docx CB1 run. On clean CB1, C-002 returns "Requires Review" (KEY DIS FAIL), and the model exhibits heavy Accept-bias. Delays performance is catastrophically weak (T2 = 35/100, CB2 = 67/100): DEL-003 misclassified, EOT outside range, no FIDIC citations in T2. Schedules generation is strong, but the SS+10 lag is misread as FS+10, producing a 55-working-day inflation in CB3. Full sub-scores on the Nemotron 4B tab.

Ministral 3 3B

Contracts weak on clean inputs (T1 = 63, CB1 = 34); the earlier "CB1 89/100" and "+26 point T→CB recovery" came from an answer-leaked docx run. Clean CB1 shows severe Accept-bias (17 of 21 clauses Accepted) and a C-002 KEY DIS FAIL. Delays performance is moderate: DEL-OW-002 (Contractor, FIDIC 4.15) is correctly classified in both tests, but the FM label is inconsistent and EOT falls outside range. Recurring invalid JSON output (// comments, arithmetic expressions in values) appears across delays tests. Schedules CPM is systematically broken: T3 circular dependencies collapse C to 0; the CB3 SS+10 misread inflates duration to 296 vs golden 265.

Granite 4 Tiny

IBM's enterprise MoE underperforms its 7B class on contracts (T1 = 33/100: four hallucinated articles, Art10 marked Accepted, KEY DIS FAIL), but shows a unique T2 strength — the first model to correctly classify DEL-003 as employer non-critical with zero EOT entitlement. Schedule generation is solid (T3 = 78/100, no circular dependencies). CB3 CPM is broken: Activity 8 is not on the critical path (SS+10 misread as FS+10, ES = 235 not 175). Very brief CB responses suggest a capacity limit at this MoE scale.

Gemma 4 E4B

Base overall 74/100. Contracts: T1 = 93/100 (all 14 articles correct, no hallucinations, Art10 Rejected ✓, strong DB IDs); CB1 = 70/100 on the clean plain-text input. The CB3 breakthrough is the first correct SS+10 lag application in the programme (ES8 = 175), achieving exact 265-working-day duration. Thinking traces (10K–12K characters) show thorough per-item reasoning. The principal weakness is T2 Delays (49/100): DEL-003 placed on the critical path, and a backward-pass error produces negative TF values (−5).

Qwen 3.5 9B

Highest overall base score (77/100). CB3 is outstanding at 96/100 — exact 265-working-day duration, perfect critical path, correct SS+10 (ES8 = 175) and SS+15, and no negative TF values (fixes Gemma's backward-pass error). Contracts is the strongest in the programme: T1 = 88/100 with Art10 Rejected and thorough 41K-character reasoning; CB1 = 89/100 on clean input. The CB1 key discriminator nonetheless fails — C-002 is labelled "Requires Review" because the model correctly identifies DB3 as "completely rejected" but argues that a negotiation-only clause does not match DB3's arbitration-specific entry, missing that Rule 3 applies to the whole Dispute Resolution category. Delays remains consistently weak (T2 = 56, CB2 = 53). T3 circular predecessor dependencies (three chains) kill the C section.

Ministral 14B

T1 contracts key discriminator passes (Art10 Rejected), but CB1 C-002 fails on the clean plain-text input (Requires Review, not Rejected; the earlier "C-002 Rejected" was an answer-leaked docx run; clean CB1 = 54/100). T2 event identification is perfect (35/35 A-section) — the best event recall in the programme. CB2 is strong (70/100): DEL-OW-002 Contractor correct, Adyard principle cited, weather EOT maintained, EOT = 35 within acceptable range. Weaknesses include T2 KEY DIS fail (DEL-003 placed on CP, EOT = 45 not 0), T3 circular dependencies 15↔16↔17 (C = 0/25), and CB3 SS+10 lag misread (ES8 = 165 not 175). Uniquely, CB3 C = 25/25 — no negative TF, cleaner backward pass than Gemma's. reasoning_content is empty across all tests and all outputs are wrapped in a markdown code fence.

GPT-OSS 20B

OpenAI's first open-weight model evaluated locally via LM Studio with reasoning_effort=low. T1 contracts key discriminator passes (Art10 Rejected, explicit Rule 3, strong T1 reasoning with Art10/7/6/3 all 6/6), but CB1 C-002 fails on the clean plain-text input (Requires Review, not Rejected). T3 has no circular dependencies — one of only three models to achieve this (with Granite and Nemotron). CB3 SS+10 is correctly applied in the forward pass (ES8 = 175), matching Gemma and Qwen, but a backward-pass error gives TF8 = 40 and Activity 8 non-critical (KEY DIS FAIL). Delays is the weakest domain (53/100): T2 DEL-003 placed on CP, EOT = 1.5 months vs golden 0; CB2 is severely truncated (29 reasoning tokens, 758 characters), missing the concurrent delay analysis and inverting the FIDIC 19.4 FM cost rule. reasoning_effort=low is insufficient for delay analysis and CPM backward-pass depth.

Ministral 3 3B — Delays

The smallest model fine-tuned in the programme (3B, non-thinking). The Delays cycle proceeds base 47 → FT 57 (+10) → FT+PE 62 (+15 vs base). FT+PE is the strongest configuration: the engineered prompt's explicit FIDIC and Adyard rule injection eliminated an invalid-JSON output defect and corrected the DEL-OW-002 contractor responsibility classification. Even a non-thinking 3B can be lifted into the mid-60s on Delays when the base has a workable foundation. Contracts and Schedules were not fine-tuned for this model; Delays was its strongest base domain (63), so the fine-tune targeted that strength.

Nemotron 3 Nano 4B — Contracts

On clean inputs the Contracts fine-tune scored 70 against base 76 — a −6 regression. The earlier "FT learned the rejection hierarchy (C-002)" claim was an artefact of an answer-leaked docx CB1 run; on clean CB1 both base and fine-tune fail C-002 (Requires Review, not Rejected). The real T1 regressions — article hallucination, over-triggered modification — persist. FT+PE recovered the score to 82, clearing base 76, which is a genuine best-of-both-worlds outcome unique to Nemotron in this cycle. The lesson for hybrid Mamba-2 architectures with limited training data is that fine-tuning alone is risky; FT + engineered prompt is the safer stack.

Gemma 4 E4B — Contracts, Delays, Schedules

The most fine-tuned model in the programme. Contracts moved from 82 to 92 (+10) — the only contracts fine-tune that improves on its base — driven by CB1 70→94 (+24); the fine-tune passes the C-002 key discriminator that the base fails and handles the Finnish YSE 1998 curveball cleanly, provided it is evaluated in thinking mode at ≥8k context. Delays moved from 62 to 75 with FT+PE (+13); this configuration returned the first correct partial EOT for the disputed TP fire in CB2 (38 days vs golden ~40). Schedules v3 moved from 79 to 84.5 with FT vanilla (+5.5); the v3 dataset rework (77 planner-pattern examples, T3+CB3 contamination eliminated) raised T3 from v2's 65 to 83 and removed hallucinations such as "Berlin". Prompt engineering actively hurt Gemma's Schedules CB3 by over-prescribing rules that broke the backward pass. The overall pattern is that Gemma absorbs the fine-tune pattern cleanly across all three domains, while PE helps Delays and hurts Schedules.

Qwen 3.5 9B — Contracts, Schedules

Contracts regressed by −11 (89→78). At temp = 0.6 the fine-tune emits the invalid string "Accepted subject to modification" on 7 of 14 T1 articles — a format defect rather than a reasoning failure. FT+PE rescues the score to 87 but still trails base 89 and base+PE 94. Schedules v3 follows a similar pattern: FT vanilla 76.5 (−2.5 vs base 79), FT+PE 73 (worse still). Crucially, Base+PE on Schedules reaches 87.5 combined (+8.5), including a perfect CB3 98/100 (exact PD = 265, exact CP, all flags consistent). The verdict for Qwen on both domains is to skip fine-tuning entirely and operate the model in Base+PE mode: strong base reasoning combined with explicit CPM rule injection produces production-grade output without any weight modification overhead.

Ministral 3 14B — Delays

The 14B Reasoning model on Delays: base 66 → FT 70 → FT+PE 79 (+19 vs base). FT+PE was the only configuration to compute the Adyard offset exactly right (42d vessel breakdown − 26d concurrent window = 16d contractor LD). The combination of 14B reasoning capacity, planner-pattern training data, and explicit FIDIC rule injection in the prompt produced the strongest Delays score in the programme. Training required the Unsloth path (raw PEFT crashed on the 4-bit Ministral3 base); per-epoch evaluation was disabled because the accelerate fp32 conversion OOMs on logits at sequence length 4096; and stream merge was mandatory for the locally-trained 4-bit adapter, since Unsloth's save_pretrained_merged corrupts on this combination.

GPT-OSS 20B — Schedules (Overall Winner)

The 20B MoE base (3.6B active) was fine-tuned on the Schedules v3 dataset on a cloud A100, the local 4090 Laptop being blocked twice (Unsloth fused-CE incompatibility on SM89, and the transformers MXFP4 hard training guard). FT+PE reached 92.5 combined (T3 = 87, CB3 = 98), the overall programme winner. Both FT vanilla and FT+PE produce a perfect CB3 98/100: PD = 265 exact, CP = [1, 2, 3, 5, 6, 7, 8, 14, 15, 17, 18] matching golden exactly, Activity 8 critical, SS+10 and SS+15 correctly applied, Activity 1 critical = true (where Gemma produced TF = −15). T3 FT+PE uses multi-DB blending (P1/P2/P4/P5/P8) with per-activity scale rationale. The complete cloud cycle cost approximately $2 and 60 minutes wall-clock; the MXFP4 GGUF (13 GB) was deployed locally for LM Studio. The pattern is the inverse of Qwen: large MoE bases combined with planner-pattern fine-tuning and an engineered prompt stack additively rather than regressing.

• Latency/throughput not measured in this phase

Overall Conclusion 1 — Best Configuration per Domain

No single pathway wins all three domains. Contracts rewards Base+PE because the task is dominated by knowledge retrieval and fine-tuning bakes in a stale snapshot of a company's clause database. Delays and Schedules reward FT+PE because the task is closed-form reasoning that transfers cleanly into fine-tuned weights. The strongest production stack selects the pathway per domain on the basis of whether the underlying task is dynamic-knowledge or closed-form-reasoning. The cross-model winners are summarised below.

Overall Conclusion 2 — Model Size Is Not the Primary Predictor

Across the 3B–20B range tested, raw parameter count showed weak correlation with final scores. The 4B Gemma E4B fine-tuned contracts (92) beats both the 14B Ministral (best contracts 66 base) and the 20B GPT-OSS (best contracts 78 base). The 9B Qwen Base+PE contracts (94) is the best contracts configuration overall. On Schedules, the 20B GPT-OSS FT+PE (92.5) does win, but Qwen 9B Base+PE (87.5) and Gemma 4B FT (84.5) are within striking distance at a fraction of the inference cost. Architecture (thinking capability, hybrid state-space, mixture-of-experts), pre-training data quality, and fine-tuning data design matter more than parameter count for construction-domain SLM performance.

Contracts — Skip Fine-Tuning, Use Base + PE + RAG Scaffolding

The three contracts fine-tunes split unevenly: Gemma 4 E4B improved (+10 combined, +24 on the CB1 Finnish YSE 1998 curveball), while Nemotron 3 Nano 4B (−6) and Qwen 3.5 9B (−11) regressed. The "trades generalisation for specialisation" pattern does not hold uniformly. Gemma's fine-tune generalised better, but Qwen's invalid-status-string format defect at temp = 0.6 and Nemotron's article hallucination wiped out fine-tuning gains.

The root cause is that contracts requires reasoning combined with retrieval against a company's evolving clause database. Fine-tuning encodes a snapshot; weights trained today are stale by next quarter as precedents shift, commercial strategy changes, and new clauses close. The architecture-correct path is a capable base (Qwen 9B Base+PE = 94 is the programme high) combined with retrieval-augmented scaffolding (clause-database injection, tool-calling, MCP) and a carefully engineered prompt. Fine-tuning bakes in what the company decided in the past; what is needed is teaching how the model reasons, and for contracts that capability is already present in the base. Fine-tuning should be reserved for the rare model–domain combination where it demonstrably clears the base; in this cycle Gemma 4 E4B contracts is the only such example.

Delays — Fine-Tuning Wins Across the Board

Three models completed the full Delays fine-tune cycle (Ministral 3 3B, Gemma 4 E4B, Ministral 3 14B), each with the same six T2-envelope training additions and the same engineered prompt. Every model improved with fine-tuning, and FT+PE was the best configuration in every case.

Delays behaves differently from Contracts because it is pure reasoning — TIA framework, float arithmetic, responsibility classification, Adyard concurrency — with no live clause database and no jurisdiction-specific knowledge that goes stale. Training data teaches how to reason, which transfers cleanly to weights. The single strongest configuration, Gemma 4 E4B FT+PE, returned the first correct partial EOT for the disputed TP fire in CB2 (38 days vs golden ~40). Ministral 14B FT+PE was the only configuration to compute the Adyard offset exactly (42d vessel breakdown minus 26d concurrent window = 16d contractor LD).

Schedules — Three Pathways All Achieve Perfect CB3, FT+PE Wins T3

The Schedules v3 cycle (planner-pattern dataset, 77 train + 11 val, T3 and CB3 task-pattern contamination eliminated) produced the strongest results of the programme. Three configurations achieved a perfect CB3 score of 98/100 on the Northbrook Solar 50MW EPC CPM analysis.

Three patterns emerged across model sizes. The small dense model (Gemma 4 E4B) prefers FT vanilla and is actively hurt by PE. The medium dense model (Qwen 3.5 9B) prefers Base+PE and regresses on CB3 under fine-tuning due to an A2 SS bug. The large MoE model (GPT-OSS 20B) shows additive stacking: both FT vanilla and FT+PE achieve perfect CB3, with PE adding +7 on T3. The v3 dataset rework — replacing v1's mixed format and v2's 8-envelope minority signal with 77 pure planner-pattern examples — was the key unlock. The planner pattern encodes activity selection, sequence logic, and duration grounding; the deterministic scaffold computes the CPM math. The model learns the correct abstraction. The CB3 Northbrook Solar critical path (1→2→3→5→6→7→8→14→15→17→18) was matched exactly by all three top configurations, demonstrating that small, medium, and large SLMs can all reach production-grade CPM analysis when paired with the right pathway.