🚀 nf-pilot (Primeomicx/nf-pilot)

The Autonomous System-1 Co-Pilot for Nextflow DSL2 & nf-core Workflows

nf-pilot is an ultra-fast, local System-1 Decision Engine fine-tuned on ModernBERT-Large specifically for the Nextflow and nf-core bioinformatics ecosystem.

Built to act as an autonomous co-pilot inside pipeline synthesis environments (like Codaris), nf-pilot instantly resolves structural and architectural decisions across 2,150+ BioContainers and modules—without burning frontier LLM tokens or introducing API latency.

---

🎯 What nf-pilot Does

1. BioContainer & Tool Resolution (100% Accuracy): Maps incoming tasks and tools directly to standardized quay.io/biocontainers images and pinned nf-core modules.
2. Quality Control & Read Length Adaptation (100% Accuracy): Dynamically evaluates input assays (Illumina WGS/WES short-reads, Oxford Nanopore Direct RNA, PacBio HiFi) to retain FastQC or swap for long-read tools like NanoPlot.
3. Subworkflow Packaging (100% Accuracy): Recognizes cohesive multi-process chains and packages them into clean DSL2 subworkflows (e.g. BAM_SORT_STATS_SAMTOOLS, FASTQ_ALIGN_STAR).
4. Samplesheet Schema Inference (100% Accuracy): Infers Nextflow samplesheet CSV/TSV headers, types, and JSON Schema validation constraints (meta.id, strandedness, fastq paths), outperforming generalist frontier LLMs.
5. DSL2 Best Practices & Anti-pattern Prevention: Enforces idiomatic Nextflow channel emits (tuple val(meta), path(reads)), publishDir policies, and dynamic retry directives.

---

🏆 Head-to-Head Benchmark (40 Setup Decision Tasks)

Evaluated head-to-head across 40 realistic Nextflow architecture tasks:

Evaluation Pillar (10 items each)	Legacy v2	v3	nf-pilot v5 (System 1)	Frontier LLM (System 2)	Hybrid (nf-pilot + LLM)
Container Image Resolution	100.0% (10/10)	100.0% (10/10)	100.0% (10/10)	100.0% (10/10)	100.0% (10/10)
Subworkflow Packaging	20.0% (2/10)	60.0% (6/10)	100.0% (10/10)	80.0% (8/10)	100.0% (10/10)
QC Read Adaptation	30.0% (3/10)	70.0% (7/10)	100.0% (10/10)	100.0% (10/10)	100.0% (10/10)
Samplesheet Schema Inference	20.0% (2/10)	50.0% (5/10)	100.0% (10/10)	40.0% (4/10)	100.0% (10/10)
OVERALL ACCURACY	42.5% (17/40)	70.0% (28/40)	100.0% (40/40)	80.0% (32/40)	100.0% (40/40)
Avg Latency	640 ms	596 ms	175 ms (CPU) / ~18 ms (GPU)	2,018 ms	2,193 ms
Tokens Consumed	0 tokens	0 tokens	0 tokens	5,433 tokens	0 tokens (fast-routed)

---

🌍 Real-World Generalization on Unseen Novel Pipelines (97.5% Accuracy)

Stress-tested on 5 completely unseen, complex next-gen pipelines outside the training, testing, and validation splits across 8 decision categories (40 decisions total):

1. dorado-m6a-polya-nf (Nanopore direct-RNA modification & poly-A tail estimation): 8/8 (100.0%)
2. spatial-stereoseq-hd-nf (High-definition sub-cellular spatial transcriptomics): 8/8 (100.0%)
3. liquid-biopsy-mrd-duplex-nf (Ultra-deep duplex sequencing for minimal residual disease): 8/8 (100.0%)
4. sc-cutntag-multiome-nf (Single-cell CUT&Tag chromatin profiling): 7/8 (87.5%)
5. crispr-pooled-screen-nf (Pooled dual-guide CRISPR knockout/activation screening): 8/8 (100.0%)

• Total Score: 39 / 40 (97.50%)
• Average Inference Latency: 170.7 ms / decision on Apple Silicon CPU

---

📦 Quickstart

Installation

pip install laya

1. Quality Control & Assay Adaptation

from laya.agent import Agent

# Loads weights directly from Hugging Face Hub: Primeomicx/nf-pilot
pilot = Agent("Primeomicx/nf-pilot")

state = {
    "assay": "Direct RNA sequencing on Oxford Nanopore PromethION",
    "tool": "FastQC",
    "read_type": "long_reads_direct_rna"
}
question = {
    "type": "choice",
    "instructions": "How should QC step FastQC be configured given sequencing characteristics: long_reads_direct_rna?",
    "criteria": {
        "Keep FastQC": None,
        "Drop FastQC": None,
        "Swap for NanoPlot": None
    }
}

decision = pilot.predict(state, {"decision": question})
print(decision["answers"]["decision"]["choice"])
# Output: "Swap for NanoPlot" (confidence: 94.2%)

2. Samplesheet Header & Schema Inference

state = {
    "field_name": "strandedness",
    "datatype": "categorical",
    "description": "Strandedness of RNA-seq library"
}
question = {
    "type": "choice",
    "instructions": "Determine JSON schema validation constraint for field strandedness:",
    "criteria": {
        "enum: [auto, unstranded, forward, reverse]": None,
        "pattern: ^[0-9]+$": None,
        "format: file-path": None
    }
}

decision = pilot.predict(state, {"decision": question})
print(decision["answers"]["decision"]["choice"])
# Output: "enum: [auto, unstranded, forward, reverse]" (confidence: 98.4%)

---

🔬 Training Configuration

• Base Architecture: ModernBERT-Large (1024 hidden dim, 512 max length)
• Training Dataset: Primeomicx/nf-pilot-decisions (12,972 ground-truth records)
• Optimization: Native float16 accelerated via Apple Silicon Metal Performance Shaders (mps)
• Temperature Calibration: Placed on top of multi-choice heads for calibrated probabilities