Trainline — Datasets

0 datasets

Trainline turns reviewed cases into LoRA training data and tracks the fine-tuned adapters built from it — the JSONL datasets below, and the trained adapters with their loss curves further down. The adapter behind the Classifier bot is trained here.

how to use this view

How to use. psyc train-build-all builds the datasets below; the Dockerfile.train workflow fine-tunes an adapter; both then appear here.

What you're seeing. JSONL datasets with example counts, and trained QLoRA adapters with their base model, hyperparameters, and per-step loss curve.

Why it matters. The adapter behind the Classifier bot is built here — this page is the provenance of the model that's actually in operation.

No datasets yet. Run psyc train-build-all.

Trainline — Adapters

1 adapter

psyc-v5 trained

Base model: unsloth/Qwen3.5-4B
Examples: 2392
Epochs: 3
LoRA r: 16
Learning rate: 0.0002
Final train loss: 0.3225
Datasets: ioc_extraction-v5.jsonl severity_classification-v5.jsonl routing_decision-v5.jsonl tlp_assignment-v5.jsonl

Training loss by step

step 10 · ep 0 2.2571

step 20 · ep 0 0.8620

step 30 · ep 0 0.6793

step 40 · ep 0 0.4463

step 50 · ep 0 0.4220

step 60 · ep 0 0.5912

step 70 · ep 0 0.4615

step 80 · ep 0 0.3874

step 90 · ep 0 0.4495

step 100 · ep 0 0.4432

step 110 · ep 0 0.3409

step 120 · ep 0 0.5136

step 130 · ep 0 0.3205

step 140 · ep 0 0.3462

step 150 · ep 0 0.2599

step 160 · ep 0 0.2541

step 170 · ep 0 0.3083

step 180 · ep 0 0.3790

step 190 · ep 0 0.3323

step 200 · ep 0 0.3296

step 210 · ep 0 0.3252

step 220 · ep 0 0.2195

step 230 · ep 0 0.3397

step 240 · ep 0 0.4363

step 250 · ep 0 0.2504

step 260 · ep 0 0.3359

step 270 · ep 0 0.3386

step 280 · ep 0 0.3956

step 290 · ep 0 0.3409

step 300 · ep 1 0.4690

step 310 · ep 1 0.1880

step 320 · ep 1 0.2629

step 330 · ep 1 0.2626

step 340 · ep 1 0.4318

step 350 · ep 1 0.3017

step 360 · ep 1 0.2637

step 370 · ep 1 0.2067

step 380 · ep 1 0.2026

step 390 · ep 1 0.3198

step 400 · ep 1 0.2455

step 410 · ep 1 0.2478

step 420 · ep 1 0.3844

step 430 · ep 1 0.4528

step 440 · ep 1 0.3035

step 450 · ep 1 0.2045

step 460 · ep 1 0.2369

step 470 · ep 1 0.2395

step 480 · ep 1 0.2657

step 490 · ep 1 0.2538

step 500 · ep 1 0.1887

step 510 · ep 1 0.1887

step 520 · ep 1 0.2400

step 530 · ep 1 0.3522

step 540 · ep 1 0.1784

step 550 · ep 1 0.3819

step 560 · ep 1 0.2669

step 570 · ep 1 0.2252

step 580 · ep 1 0.3038

step 590 · ep 1 0.3403

step 600 · ep 2 0.2664

step 610 · ep 2 0.3956

step 620 · ep 2 0.3756

step 630 · ep 2 0.1365

step 640 · ep 2 0.2091

step 650 · ep 2 0.2396

step 660 · ep 2 0.2061

step 670 · ep 2 0.3007

step 680 · ep 2 0.3434

step 690 · ep 2 0.2620

step 700 · ep 2 0.2732

step 710 · ep 2 0.2007

step 720 · ep 2 0.2418

step 730 · ep 2 0.1590

step 740 · ep 2 0.1114

step 750 · ep 2 0.1654

step 760 · ep 2 0.1623

step 770 · ep 2 0.2910

step 780 · ep 2 0.1641

step 790 · ep 2 0.2017

step 800 · ep 2 0.2143

step 810 · ep 2 0.1503

step 820 · ep 2 0.2297

step 830 · ep 2 0.2295

step 840 · ep 2 0.2496

step 850 · ep 2 0.2145

step 860 · ep 2 0.2374

step 870 · ep 2 0.2963

step 880 · ep 2 0.2293

step 890 · ep 2 0.1505