Fine-Tuning Pipelines
Managed SFT · LoRA · RLHF · Distillation — reproducible on your data
job_a7c1 — cortex-med-7b SFT round 12
8 × H100 · us-east-2 · started 4h 22m ago
Step
6,800 / 10,000
Tokens/s
42,180
Loss
0.284
Learning rate
3.2e-5
Grad norm
0.42
Perplexity
1.34
ETA
1h 12m
Cost so far
$284.10
Live logs
stdout · streaming
[14:22:11] rank=0 step=6800 loss=0.284 lr=3.2e-5 [14:22:12] rank=1 step=6800 loss=0.281 lr=3.2e-5 [14:22:12] rank=2 step=6800 loss=0.286 lr=3.2e-5 [14:22:13] rank=3 step=6800 loss=0.283 lr=3.2e-5 [14:22:14] eval: perplexity=1.34 acc=0.941 [14:22:15] checkpoint saved: s3://cortex-med/ckpt-6800 [14:22:18] tokens/s=42,180 gpu_util=94% [14:22:20] grad_norm=0.42 clip=1.0 [14:22:21] rank=0 step=6801 loss=0.279 lr=3.2e-5 [14:22:22] rank=1 step=6801 loss=0.282 lr=3.2e-5 [14:22:23] enclave attestation refreshed ok [14:22:25] rank=0 step=6802 loss=0.277 lr=3.2e-5 [14:22:26] rank=1 step=6802 loss=0.280 lr=3.2e-5 [14:22:28] warmup complete after 500 steps [14:22:30] tokens/s=42,240 gpu_util=95%
autoscroll: onlines: 12,482
Job queue
All fine-tuning jobs across workspace
job_a7c1cortex-med-7b — SFT round 12
James Okafor
8 × H100· us-east-2· ETA 1h 12m· $284.10
68%Running
job_b920radiology-vit — LoRA sweep
Anna Weiss
4 × A100· eu-west-1· ETA 3h 04m· $96.80
42%Running
job_c412fraud-detector — retrain
Priya Mehta
2 × A10· us-west-2· ETA -· $0.00
0%Queued
job_d001voice-triage — RLHF
Sara Kim
16 × H100· us-east-2· ETA 18m· $1,204.22
91%Running
job_e552claims-nlp — domain adapt
Marcus Lin
4 × A100· eu-west-1· ETA -· $52.40
33%Failed
job_f118notes-summarizer — SFT
James Okafor
8 × H100· us-east-2· ETA -· $612.00
100%Completed
Recipe · SFT + LoRA
recipe: sft-lora-v3 base_model: llama-3.1-8b method: lora lora_rank: 16 lora_alpha: 32 target_modules: [q_proj, k_proj, v_proj, o_proj] optimizer: adamw learning_rate: 3.2e-5 warmup_steps: 500 batch_size: 32 gradient_accumulation: 4 max_steps: 10000 eval_every: 200 save_every: 500 enclave: intel_tdx dataset: readmit-notes-2025q1@v14
Hyperparameter sweep
lr=1e-5
0.312
lr=3e-5
0.289
lr=5e-5
0.284
best
lr=1e-4
0.302
lr=3e-4
0.358