Primeomicx/nfcore-laya-decisions

nf-core & Nextflow Laya Decision Engine (laya-nfcore-v1)

A standardized decision-layer dataset and fine-tuned model for Laya (System-1 Decision Engines) across Nextflow DSL2 and the complete nf-core bioinformatics ecosystem (2,153 modules, 101 pipelines).

🏆 Test Set Performance (691 Held-Out Records)

Evaluating laya-nfcore-v1 on the independent held-out test split (data/test.jsonl):

Evaluation Category	Test Set Score	Accuracy	Description
DSL2 Syntax & Rules (noul)	165 / 165	100.0%	Full resolution of Nextflow idioms, channel emits, and anti-patterns
Tool & Module Selection	421 / 447	94.18%	Disambiguating bioinformatics tools across 2,153 modules
User Intent Routing	27 / 29	93.10%	Routing queries to code gen, debugging, config, or indexing
nf-core Pipeline Routing	45 / 50	90.00%	Matching assay requirements to official nf-core workflows
OVERALL ACCURACY	658 / 691	95.22%	Inference Latency: ~45–65 ms on Apple Silicon

---

📊 Dataset Summary

• Total Decision Records: 6,911
• Domain Coverage:
  • Tool & Module Selection (4,306 records): Spans all 2,153 nf-core modules and subworkflows with mined hard-negative distractors (aligners, variant callers, QC, trimming, single-cell, assembly, etc.).
  • Pipeline Routing (505 records): Covers all 101 official nf-core pipelines across diverse assay queries (RNA-seq, Sarek variant calling, Ampliseq, ATAC-seq, Viralrecon, etc.).
  • DSL2 Syntax & Rules Check (1,800 records): Balanced binary noul decisions evaluating idiomatic DSL2 patterns (tuple val(meta), path(reads), publishDir, take/main/emit) versus common anti-patterns.
  • Task & Intent Routing (300 records): Intent classification across pipeline execution, error debugging, custom module design, and reference genome preparation.

🗂️ Split Information

Split	Records	File
Train	5,529	data/train.jsonl
Validation	691	data/validation.jsonl
Test	691	data/test.jsonl

📁 Repository Structure

laya_training/
├── README.md                      # Hugging Face Dataset Card & Documentation
├── requirements.txt               # Dependencies
├── inputs/                        # Raw source corpora
│   ├── nfcore_corpus.json         # Complete 2,153 nf-core module corpus
│   └── nfcore_pipeline_catalog.json # Complete 101 nf-core pipeline catalog
├── data/                          # Hugging Face standard dataset format
│   ├── train.jsonl                # 5,529 training records
│   ├── validation.jsonl           # 691 validation records
│   ├── test.jsonl                 # 691 test records
│   ├── cached_train_h.pt          # Pinned ModernBERT hidden state cache
│   ├── cached_val_h.pt            # Validation hidden state cache
│   └── dataset_dict.json          # Dataset metadata
├── scripts/                       # Training, evaluation & upload pipeline
│   ├── 01_build_dataset.py        # Generates JSONL splits from corpora
│   ├── 02_train_laya.py           # RAM-cached ModernBERT decision trainer
│   ├── 03_evaluate.py             # 100-question Nextflow benchmark runner
│   └── 04_push_to_hub.py          # Hugging Face Hub upload utility
├── checkpoints/                   # Checkpoints ready for HF Hub & Laya Agent
│   └── laya-nfcore-v1/
│       ├── model.safetensors      # Tuned weights (1.6 GB)
│       ├── rl_agent_config.json   # Laya config
│       ├── encoder/               # ModernBERT configuration
│       └── tokenizer/             # Tokenizer vocabulary & configs
└── results/                       # Benchmark reports & training logs
    ├── baseline_benchmark_100.json
    ├── post_training_benchmark_100.json
    ├── test_set_evaluation.json
    └── training_history.json

📋 Data Schema

Each row in data/*.jsonl is formatted for direct consumption by Laya agents and Hugging Face datasets:

{
  "id": "mod_angsd_realsfs_1",
  "category": "tool_selection",
  "state": "Which module should be imported to execute angsd for Estimate site frequency spectrum from site allele frequencies?",
  "question": {
    "type": "choice",
    "instructions": "Select the appropriate nf-core module for the specified bioinformatics operation.",
    "criteria": {
      "angsd/dosaf": "Estimate site allele frequencies from BAM files.",
      "angsd/realsfs": "Estimate site frequency spectrum from site allele frequencies",
      "cnvkit/access": "Calculate the sequence-accessible coordinates in chromosomes from the given reference genome",
      "mm2plus/index": "Provides fasta index required by mm2plus alignment."
    }
  },
  "target": "angsd/realsfs",
  "target_idx": 1
}

🚀 Quickstart

1. Load via Hugging Face datasets

from datasets import load_dataset

dataset = load_dataset("json", data_files={
    "train": "data/train.jsonl",
    "validation": "data/validation.jsonl",
    "test": "data/test.jsonl"
})
print(dataset)

2. Fine-Tune Laya

python scripts/02_train_laya.py --epochs 8 --batch_size 32 --lr 6e-5 --device mps

3. Evaluate Model

python scripts/03_evaluate.py --model_path checkpoints/laya-nfcore-v1

4. Push to Hugging Face Hub

python scripts/04_push_to_hub.py --dataset-repo <your-username>/nfcore-laya-decisions --model-repo <your-username>/laya-nextflow-nfcore