Finetune Status | Zyeon UI

Preview in your theme

job.status='training' — a live run. Scrub the loss curve with the pointer or the arrow keys; Cancel opens an inline destructive confirm, and the callback fires once per run no matter how many times you hit it.

llama-3.1-8b-instruct · support-tone-v3

ftjob-7Hn4pXsdR9wE

Training

Validating— done
Queued— done
Training— in progress
Result— not started

212 of 480 steps · epoch 1.3 of 344%

4.8M tokens trained

Loss

trainvalidation21 samples · train 1.955 → 0.573 · best val 0.602 at step 210

Hyperparameters

N epochs: 3
Batch size: 16
Learning rate: 2.0e-5
Lr scheduler: cosine
Warmup ratio: 0.03
Lora rank: 16
Lora alpha: 32
Max seq len: 4096
Precision: bf16
Packing: on

Checkpoints· 1

no callback fired yet

job.status='succeeded' — 12 checkpoints capped at 4 with an explicit 'Show all 12'. Selecting a row is what re-targets 'Use this model'; the dashed rule in the chart marks the best held-out loss (step 360), which is NOT the last step — held-out loss climbs from 0.574 to 0.781 while training loss keeps falling to 0.441.

llama-3.1-8b-instruct · support-tone-v3

ftjob-3Qa9vLbmC4pR

Succeeded

Validating— done
Queued— done
Training— done
Succeeded— done

480 of 480 steps · epoch 3.0 of 3100%

11.4M tokens trained

Loss

trainvalidation48 samples · train 1.929 → 0.441 · best val 0.574 at step 370

Hyperparameters

N epochs: 3
Batch size: 16
Learning rate: 2.0e-5
Lr scheduler: cosine
Warmup ratio: 0.03
Lora rank: 16
Lora alpha: 32
Max seq len: 4096
Precision: bf16
Packing: on

Checkpoints· 12

ft:llama-3.1-8b-instruct:acme:support-tone-v3:9dQ2

nothing selected — Use this model would ship the run's final model

job.status='failed' — the run kept the stage it died at ('while training'), so the strip shows validating/queued done and the outcome red. The CUDA trace is rendered whole: no max-height, no clamp.

llama-3.1-8b-instruct · support-tone-v4

ftjob-5Zt1kWyuN6bH

Failed

Validating— done
Queued— done
Training— failed here
Failed— failed here

96 of 480 steps · epoch 0.6 of 320%

2.2M tokens trained

Failure · while training

RuntimeError: CUDA out of memory. Tried to allocate 2.10 GiB.
GPU 0 — 79.15 GiB capacity · 74.02 GiB already allocated · 1.84 GiB free.

  at step 96 of 480 (micro-batch 3 of 4, sequence length 8192)
  at torch/nn/functional.py:2515 in scaled_dot_product_attention
  at trainer/loop.py:311 in _forward_backward
  at trainer/loop.py:184 in train

Hint: batch_size 32 at max_seq_len 8192 needs ~86 GiB. Halve the batch size,
or turn on gradient checkpointing and retry — the dataset itself validated fine.

Loss

trainvalidation9 samples · train 1.985 → 0.958 · best val 0.969 at step 90

Hyperparameters

N epochs: 3
Batch size: 32
Learning rate: 2.0e-5
Lr scheduler: cosine
Warmup ratio: 0.03
Lora rank: 16
Lora alpha: 32
Max seq len: 8192
Precision: bf16
Packing: on

Checkpoints

This run never wrote a checkpoint.

job.status='validating' — nothing has run, so the bar is indeterminate rather than claiming 0%, the curve says why it is empty, and the checkpoint list explains when the first snapshot lands.

llama-3.1-8b-instruct · support-tone-v3

ftjob-8Kd2mQrvT1aZ

Validating

Validating— in progress
Queued— not started
Training— not started
Result— not started

0 of 480 steps · epoch 0.0 of 3—

Loss

No samples yet — the curve starts once the first optimizer steps report.

Hyperparameters

N epochs: 3
Batch size: 16
Learning rate: 2.0e-5
Lr scheduler: cosine
Warmup ratio: 0.03
Lora rank: 16
Lora alpha: 32
Max seq len: 4096
Precision: bf16
Packing: on

Checkpoints

The first snapshot is written at the end of epoch 1.

A one-sample curve with no eval yet — the geometry edge that makes a naive sparkline divide by zero. It draws a dot, keeps the readout truthful, and the 'val' reading says 'not measured' instead of inventing a number.

mistral-7b-instruct · extraction-pilot

ftjob-1Bv6cFqoJ2sD

Training

Validating— done
Queued— done
Training— in progress
Result— not started

10 of 60 steps · epoch 0.2 of 117%

96K tokens trained

Loss

trainvalidation1 sample · train 2.104

Hyperparameters

N epochs: 1
Batch size: 4
Learning rate: 5.0e-5

Checkpoints

The first snapshot is written at the end of epoch 1.

status='loading' — the status endpoint hasn't answered yet, so the card is a skeleton with aria-busy and no invented numbers.

status='empty' — this project has never started a run. Not an error, and not a zero-progress card.

No fine-tune run yet

Upload a training file and start a run — its progress, loss curve and checkpoints appear here.

status='error' — the STATUS ENDPOINT failed, which is a different failure from the training run failing, so it reads differently and offers a retry.

Couldn't load this run

The training API returned 503 while polling this run. request-id: 5b81-c40e

Loading preview…

This component is part of Pro

The install command, prompt and concept map are included with a Pro license — one purchase unlocks every Pro component, forever.

Unlock with Pro

RuntimeError: CUDA out of memory. Tried to allocate 2.10 GiB. GPU 0 — 79.15 GiB capacity · 74.02 GiB already allocated · 1.84 GiB free. at step 96 of 480 (micro-batch 3 of 4, sequence length 8192) at torch/nn/functional.py:2515 in scaled_dot_product_attention at trainer/loop.py:311 in _forward_backward at trainer/loop.py:184 in train Hint: batch_size 32 at max_seq_len 8192 needs ~86 GiB. Halve the batch size, or turn on gradient checkpointing and retry — the dataset itself validated fine.