Dataset Viewer

The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

PLC-LLM-Bench (v0.2)

An offline functional, compiler, and dialect benchmark evaluating open-weight LLMs on Siemens TIA Portal SCL and IEC 61131-3 Structured Text.

plc-llm-bench measures whether local models running on consumer/workstation GPUs (starting with an NVIDIA GeForce RTX 5070 Ti Mobile, 12 GB VRAM) can generate industrial automation FUNCTION_BLOCK logic that:

  1. Follows Siemens TIA Portal SCL conventions (#var prefixing, { pragmas }, double-quoted block names, standard IEC timers/counters),
  2. Transpiles cleanly (scl_to_iec) and compiles natively with MatIEC (iec2c) + gcc (zero syntax errors, plus 1-turn Pass@2 compiler self-repair), and
  3. Passes deterministic multi-scan PLC simulation unit tests (SCLFunctionBlock) across timers, edge triggers, PID loops, grid droop curves, ring buffers, arrays, motion homing, Modbus CRC-16, and 2oo3 voting.

PLC-LLM-Bench Pareto & Accuracy Charts


1. Three-Layer Verification Pipeline

  1. Layer 1 — TIA Portal SCL Dialect Linter (plc_bench.scl_normalizer.lint_scl) Scores structural adherence to Siemens TIA Portal SCL: FUNCTION_BLOCK / FUNCTION headers, double-quoted block identifiers, { S7_Optimized_Access := 'TRUE' } / VERSION, #var local symbol references, and balanced VAR*/END_VAR and block terminators.
  2. Layer 2 — v0.2 Transpiler & Native Compiler (plc_bench.scl_normalizer.scl_to_iec + plc_bench.compiler.compile_scl) Lowers TIA Portal SCL sugar into strict IEC 61131-3 Structured Text (BEGIN stripping with trailing comments preserved, // comment stripping outside string literals, named built-in arguments such as LIMIT(MN:=..., IN:=..., MX:=...), => FB output extraction, VAR_STAT / IEC_TIMER normalization, and Real := int; coercion), then compiles through MatIEC iec2c -> C -> gcc. If Pass@1 compilation fails, the exact compiler diagnostics are fed back for a 1-turn Pass@2 repair attempt.
  3. Layer 3 — Multi-Scan Deterministic PLC Simulator (plc_bench.scl_interp.SCLFunctionBlock) Executes the generated FUNCTION_BLOCK across dozens/hundreds of PLC scan cycles with deterministic scan_dt_ms time steps, simulating TON/TOF/TP/IEC_TIMER, R_TRIG/F_TRIG, CTU/CTD, SR/RS, ARRAY[lo..hi], VAR_IN_OUT, bitwise masks / .%X bit access, EXPT/ROL/ROR, CONTINUE/EXIT/RETURN, and control flow.

2. Benchmark Dataset & Task Catalog (30 Custom SCL + Academic Subset)

The exported dataset in dataset/plc_scl_tasks.jsonl contains 30 verified multi-scan Siemens SCL tasks across 5 industrial categories (6 tasks per category). All 30 reference solutions pass Layer 1 linting, native MatIEC iec2c + gcc compilation, and Layer 3 multi-scan assertions (tests/test_tasks.py).

Custom Siemens SCL Tasks (plc_bench/tasks_scl.py)

Category Task IDs Function Blocks Tested Key Capabilities Verified
Timers & State Machines SCL-SM-01 .. SCL-SM-06 FB_MotorDOL, FB_StarDelta, FB_ConveyorJam, FB_PumpLeadLag, FB_BatchSequencer, FB_PackML_State Seal-in interlocks, TON star-delta dead-time, jam watchdog, edge-triggered lead/lag alternation, timed batch sequencer, PackML state transitions (EXECUTE/HOLDING/HELD/STOPPING/STOPPED)
Process & Analog Control SCL-PC-01 .. SCL-PC-06 FB_Scale4to20mA, FB_LowPassEMA, FB_PID_AntiWindup, FB_RampLimiter, FB_SplitRange, FB_AnalogAlarmHyst Linear scaling + NAMUR NE43 wire-break fault, EMA filter, discrete PI/PID with conditional anti-windup, rate/slew limiter, heating/cooling split-range, HH/H/L/LL alarm hysteresis
Energy, BESS & Grid SCL-EN-01 .. SCL-EN-06 FB_LFSM_Droop, FB_BESS_SafetyGuard, FB_Synchrocheck, FB_ModbusMap, FB_PV_BESS_Alloc, FB_QV_VoltVarDroop Frequency-watt LFSM-O/U droop, BESS SoC/cell-voltage/thermal derating + fault latch, grid synchrocheck window timer, IEEE-754/scaled Modbus word packing, PV/BESS dispatch allocation, Q(V) volt-var droop
Debug & Refactoring SCL-BF-01 .. SCL-BF-06 FB_PulseCounterFix, FB_SafeLatchFix, FB_LadderToSCL, FB_FlowRatioSafe, FB_EnablePIFix, FB_DebounceTimerFix Double-counting bug fix via R_TRIG, unsafe reset priority fix (Stop/E-stop overrides Start), ladder-to-SCL refactor, divide-by-zero/clamp hardening, integrator reset on disable, TON-based contact debounce
Advanced Arrays, Motion & Voting SCL-AD-01 .. SCL-AD-06 FB_MovingAvg4, FB_LeadLag3Pump, FB_AxisHomingSeq, FB_CRC16_ModbusWord, FB_Voter2oo3, FB_CascadeLoop 4-sample ARRAY ring buffer average, runtime-balanced 3-pump selector, multi-step axis homing (WAIT_HOME/BACKOFF/DONE/FAULT), bitwise Modbus CRC-16 (0xA001), 2-out-of-3 analog sensor median voting with discrepancy alarm, outer-to-inner cascade loop

Academic IEC 61131-3 Compile Subset

When cloned via --autoplc-dir and --agents4plc-dir, plc_bench/tasks_academic.py additionally evaluates models on 16 academic ST tasks sourced from AutoPLC and Agents4PLC_release using MatIEC iec2c + gcc compilation.


3. Empirical Leaderboard (v0.2 Rescored — 468 Colab T4 Evaluations)

All 13 models (Tier 1: 8 models <=12 GB VRAM; Tier 2: 5 models 12–16 GB / MoE) were executed on a Google Colab Tesla T4 (15,360 MiB VRAM) using llama-server (CUDA, q8_0 KV cache) over the 36-task v0.1 suite (20 custom SCL + 16 academic ST = 468 empirical evaluations) and re-scored cleanly with the v0.2 transpiler & simulator (rescore_checkpoint):

  • Composite Score Formula (plc_bench/report.py): $$\text{Composite} = 0.45 \cdot \text{Func P@1} + 0.25 \cdot \text{Func P@2} + 0.15 \cdot \text{Custom Compile P@1} + 0.10 \cdot \text{SCL Lint} + 0.05 \cdot \min(\text{tok/s}, 60)/60 \times 100$$
Rank Model Tier Quant Func Pass@1 Func Pass@2 Compile P@1 Compile P@2 SCL Lint Gen tok/s Peak VRAM RTX 5070 Ti Mobile Fit Composite
1 gemma-4-26b-a4b T2 gemma-4-26B-A4B-it-UD-Q3_K_M.gguf 65.0% 85.0% 50.0% 80.0% 71.2% 35.6 12.37 GB 12-16 GB / MoE Offload 68.1
2 qwen3.5-9b T1 Qwen3.5-9B-Q5_K_M.gguf 35.0% 60.0% 45.0% 80.0% 38.6% 33.5 6.00 GB 100% VRAM (<=12 GB) 44.1
3 seed-coder-8b T1 Seed-Coder-8B-Instruct-Q6_K.gguf 30.0% 30.0% 35.0% 35.0% 91.6% 33.9 6.71 GB 100% VRAM (<=12 GB) 38.2
4 qwen3-coder-30b-a3b T2 Qwen3-Coder-30B-A3B-Instruct-Q3_K_S.gguf 25.0% 25.0% 50.0% 70.0% 84.8% 49.1 12.83 GB 12-16 GB / MoE Offload 37.6
5 granite-4.0-h-small T2 granite-4.0-h-small-Q3_K_S.gguf 25.0% 25.0% 40.0% 50.0% 87.8% 21.1 13.69 GB 12-16 GB / MoE Offload 34.0
6 gpt-oss-20b T2 gpt-oss-20b-MXFP4.gguf 20.0% 30.0% 40.0% 55.0% 67.8% 44.3 11.02 GB 100% VRAM (<=12 GB) 33.0
7 devstral-small-2-24b T2 Devstral-Small-2-24B-Instruct-2512-IQ4_XS.gguf 15.0% 20.0% 45.0% 65.0% 77.8% 13.8 12.52 GB 12-16 GB / MoE Offload 27.4
8 gemma-4-e4b T1 gemma-4-E4B-it-Q4_K_M.gguf 15.0% 15.0% 15.0% 25.0% 95.4% 50.9 3.30 GB 100% VRAM (<=12 GB) 26.5
9 qwen3-14b T1 Qwen3-14B-Q4_K_M.gguf 10.0% 25.0% 35.0% 75.0% 71.2% 24.8 9.02 GB 100% VRAM (<=12 GB) 25.2
10 qwen2.5-coder-14b T1 qwen2.5-coder-14b-instruct-q4_k_m.gguf 15.0% 15.0% 30.0% 40.0% 75.1% 24.7 9.12 GB 100% VRAM (<=12 GB) 24.6
11 gemma-4-12b T1 gemma-4-12b-it-Q4_K_M.gguf 10.0% 30.0% 20.0% 45.0% 32.8% 26.7 7.32 GB 100% VRAM (<=12 GB) 20.5
12 deepseek-r1-distill-14b T1 DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf 5.0% 5.0% 20.0% 25.0% 72.2% 24.5 9.12 GB 100% VRAM (<=12 GB) 15.8
13 phi-4-reasoning-14b T1 Phi-4-reasoning-Q4_K_M.gguf 0.0% 10.0% 20.0% 25.0% 52.9% 25.2 9.32 GB 100% VRAM (<=12 GB) 12.9

Academic Dataset Subset (AutoPLC & Agents4PLC MatIEC Compile Check)

Model Tasks Academic Compile P@1 Academic Compile P@2
gemma-4-26b-a4b 16 18.8% 18.8%
qwen3.5-9b 16 6.2% 25.0%
seed-coder-8b 16 6.2% 6.2%
qwen3-coder-30b-a3b 16 31.2% 31.2%
granite-4.0-h-small 16 25.0% 25.0%
gpt-oss-20b 16 12.5% 12.5%
devstral-small-2-24b 16 12.5% 18.8%
gemma-4-e4b 16 12.5% 18.8%
qwen3-14b 16 31.2% 31.2%
qwen2.5-coder-14b 16 18.8% 18.8%
gemma-4-12b 16 12.5% 12.5%
deepseek-r1-distill-14b 16 6.2% 12.5%
phi-4-reasoning-14b 16 0.0% 6.2%

4. Hardware Guidance (RTX 5070 Ti Mobile 12 GB vs. 16 GB+ / MoE)

  • Best Strictly <=12 GB Mobile GPU Fit (qwen3.5-9b, Q5_K_M): Uses only 6.00 GB VRAM (leaving 6 GB headroom for TIA Portal / IDE), achieves 60.0% Func Pass@2, 80.0% Custom Compile P@2, and 33.5 tok/s.
  • Highest Overall Accuracy (gemma-4-26b-a4b, UD-Q3_K_M): 65.0% Func Pass@1, 85.0% Func Pass@2, 80.0% Custom Compile P@2, and 35.6 tok/s at 12.37 GB peak VRAM (3.8B active parameters MoE). Fits 16 GB desktop/laptop GPUs cleanly or runs on 12 GB GPUs with minimal MoE expert CPU offload (-ot).
  • Best Sub-12 GB Dialect Adherence (seed-coder-8b, Q6_K): 91.6% SCL lint score, 30.0% Func Pass@1, 6.71 GB VRAM, 33.9 tok/s.
  • Ultra-Fast Lightweight Draft Model (gemma-4-e4b, Q4_K_M): 50.9 tok/s, 95.4% SCL lint score, 3.30 GB VRAM.
  • MoE vs. Dense Speed Advantage: Active-parameter MoEs (qwen3-coder-30b-a3b at 49.1 tok/s, gpt-oss-20b at 44.3 tok/s, gemma-4-26b-a4b at 35.6 tok/s) outpace dense 24B (devstral-small-2-24b at 13.8 tok/s) by 2.6x–3.6x.

5. Installation & CLI Usage

Setup & Verification

pip install -e ".[dev]"
pytest -q

(Install MatIEC iec2c + gcc on Linux/WSL/Colab/Kaggle for native Layer 2 IEC 61131-3 compilation.)

Run Local GGUF Roster (llama-server)

# Quick smoke test on 1 model and 3 custom SCL tasks
python scripts/run_bench.py --smoke --models qwen3.5-9b

# Evaluate Tier 1 (<=12 GB) on all 30 custom SCL tasks
python scripts/run_bench.py --tier 1 --custom-only

# Evaluate Tier 2 (12–16 GB / MoE)
python scripts/run_bench.py --tier 2 --custom-only

Benchmark Any External OpenAI-Compatible Endpoint (Ollama, vLLM, LM Studio)

# Ollama example
python scripts/run_bench.py \
  --api-url http://localhost:11434/v1 \
  --api-model qwen3:14b \
  --api-tier 1 \
  --custom-only

# vLLM / LM Studio with reasoning prompt hint
python scripts/run_bench.py \
  --api-url http://localhost:8000/v1 \
  --api-model gemma-4-26b-a4b \
  --api-tier 2 \
  --api-reasoning \
  --custom-only

Re-Score Existing Checkpoints Offline

# Via standalone wrapper (writes .jsonl.bak backup, updates checkpoint, regenerates reports)
python scripts/rescore_checkpoint.py results/checkpoint.jsonl --out-dir results

# Or via run_bench.py
python scripts/run_bench.py --rescore-only --checkpoint results/checkpoint.jsonl --out-dir results

Export Dataset & Regenerate Colab/Kaggle Notebook

python scripts/export_dataset.py
python scripts/make_notebook.py

6. Publishing to Hugging Face & Kaggle

Hugging Face Datasets

  1. Ensure dataset/ is freshly exported and README.md (with the YAML Dataset Card above) is at the repository root:
    pip install -U huggingface_hub
    hf auth login
    python scripts/export_dataset.py
    hf repos create iulio/plc-llm-bench --type dataset --exist-ok
    hf upload iulio/plc-llm-bench . . --repo-type dataset --exclude ".venv/**" --exclude "cell*" --exclude ".*_cache/**" --exclude "**/__pycache__/**" --exclude "*.egg-info/**"
    

Kaggle Datasets & Notebooks

  1. Configure ~/.kaggle/kaggle.json (or KAGGLE_API_TOKEN) for "id": "iulianamaricai/siemens-scl-iec61131-benchmark" in dataset/dataset-metadata.json, then publish:
    pip install -U kaggle
    kaggle datasets create -p dataset
    
  2. Upload PLC_LLM_Bench_Colab.ipynb as a Kaggle Notebook (Accelerator: GPU T4 x2 or P100, Internet: On). The notebook automatically detects /kaggle/working vs. /content and runs the full MatIEC + llama-server pipeline end-to-end.

7. License, Attribution & Commercial Use

  • Dataset & Benchmark License: Released under CC-BY-NC-SA-4.0 (Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International) — see LICENSE.
  • Attribution: Academic researchers, students, and non-commercial practitioners may freely use, benchmark, and build upon this dataset with attribution to Iulian Amăricăi / CTRL ALT DELIGHTED S.R.L. (PLC-LLM-Bench).
  • Commercial Licensing: Commercial use — including training or fine-tuning proprietary commercial LLMs, commercial product evaluation, or custom enterprise industrial SCL / IEC 61131-3 / AAS Digital Twin benchmark suites — requires a separate commercial license from CTRL ALT DELIGHTED S.R.L. (Romania; creators of TwinForge AI and NexusControl Studio).
Downloads last month
76