Run Comparison

Preview in your theme

ready — 5 runs × 10 hyperparameters (4 identical) × 5 metrics, baseline sft-baseline-v3. Click a column's Set baseline to re-measure every delta against it, click a metric row heading to reorder the columns by it (best → worst → payload order), and use Hide identical parameters to drop the four rows that never change.

5 runs10 hyperparameters, 6 varying5 metricsbaseline sft-baseline-v3

Run comparison. 5 runs compared across 10 hyperparameters and 5 metrics, measured against baseline sft-baseline-v3.
Run	sft-baseline-v3main · 4a91c2e	lr-sweep-3e-4exp/lr-sweep · 8d1f077	lr-sweep-5e-5exp/lr-sweep · 8d1f077	long-context-8kexp/long-ctx · c72b410	batch-64-ckptexp/throughput · e05d9aa
Hyperparameters	6 of 10 differ between these runs
Learning ratevaries	0.0001	0.0003	0.00005	0.0001	0.0001
Batch size (per device)varies	32	32	32	16	64
Epochs	3	3	3	3	3
Optimizer	adamw	adamw	adamw	adamw	adamw
LR schedulevariesApplied after warmup	cosine	cosine	cosine	linear	cosine
Warmup ratio	0.03	0.03	0.03	0.03	0.03
Adam epsilon	1e-8	1e-8	1e-8	1e-8	1e-8
Max sequence length (tokens)varies	4,096	4,096	4,096	8,192	4,096
Gradient checkpointingvariesTrades compute for memory	false	false	false	true	true
Random seedvariesNot logged before the tracker upgrade	1,337	1,337	1,337	1,337	Not recorded
Metrics	Best value marked per row; deltas are against the baseline
	89.41%	91.28%+1.87 pp · +2.1%	88.02%-1.39 pp · -1.6%	91.35%best+1.94 pp · +2.2%	90.10%+0.69 pp · +0.8%
	0.412	0.361-0.051 · -12.4%	0.448+0.036 · +8.7%	0.357best-0.055 · -13.3%	0.389-0.023 · -5.6%
	1,840	1,795-45 · -2.4%	1,852+12 · +0.7%	1,120-720 · -39.1%	2,310best+470 · +25.5%
	78 min	79 min+1 min · +1.3%	77 min-1 min · -1.3%	154 min+1.3 h · +97.4%	52 minbest-26 min · -33.3%
	38.4	38.6+0.2 · +0.5%	38.2best-0.2 · -0.5%	71.9+33.5 · +87.2%	Not recorded

loading — skeleton with the real column geometry; no sort or baseline controls, because a focus stop inside an aria-hidden subtree has no name.

Run comparison

empty — status="empty"; the same panel also catches a ready payload with zero runs or zero rows.

Run comparison

No runs to compare

Select at least two runs — one on its own has nothing to be measured against.

error — onRetry is wired: pressing Try again swaps this instance to the ready payload.

Run comparison

Couldn't load these runs

Nothing is shown rather than a partial table — a run missing from a comparison reads as a run that lost.

Ready · sorted by validation loss, identical rows folded — defaultSort picks the best loss first, which on a lower-is-better row means ascending; defaultHideIdentical starts with only the varying parameters on screen. Peak GPU memory is missing for one run, so that column sinks to the end when you sort by it and is never marked best.

Fine-tuning sweep · week 32

Five fine-tuning runs on the same 42k-sample instruction set.

5 runs10 hyperparameters, 6 varying5 metricsbaseline sft-baseline-v3

Fine-tuning sweep · week 32. 5 runs compared across 10 hyperparameters and 5 metrics, measured against baseline sft-baseline-v3.
Run	sft-baseline-v3main · 4a91c2e	long-context-8kexp/long-ctx · c72b410	lr-sweep-3e-4exp/lr-sweep · 8d1f077	batch-64-ckptexp/throughput · e05d9aa	lr-sweep-5e-5exp/lr-sweep · 8d1f077
Hyperparameters	6 of 10 differ between these runs
Learning ratevaries	0.0001	0.0001	0.0003	0.0001	0.00005
Batch size (per device)varies	32	16	32	64	32
LR schedulevariesApplied after warmup	cosine	linear	cosine	cosine	cosine
Max sequence length (tokens)varies	4,096	8,192	4,096	4,096	4,096
Gradient checkpointingvariesTrades compute for memory	false	true	false	true	false
Random seedvariesNot logged before the tracker upgrade	1,337	1,337	1,337	Not recorded	1,337
Metrics	Best value marked per row; deltas are against the baseline
	89.41%	91.35%best+1.94 pp · +2.2%	91.28%+1.87 pp · +2.1%	90.10%+0.69 pp · +0.8%	88.02%-1.39 pp · -1.6%
	0.412	0.357best-0.055 · -13.3%	0.361-0.051 · -12.4%	0.389-0.023 · -5.6%	0.448+0.036 · +8.7%
	1,840	1,120-720 · -39.1%	1,795-45 · -2.4%	2,310best+470 · +25.5%	1,852+12 · +0.7%
	78 min	154 min+1.3 h · +97.4%	79 min+1 min · +1.3%	52 minbest-26 min · -33.3%	77 min-1 min · -1.3%
	38.4	71.9+33.5 · +87.2%	38.6+0.2 · +0.5%	Not recorded	38.2best-0.2 · -0.5%

Ready · 8 runs, no baseline — all eight columns render (the table scrolls sideways, the parameter column stays put); baselineRunId is null, so best values are still marked and no deltas are shown, because there is nothing to subtract from.

Grid sweep · 2×2×2

8 runs5 hyperparameters, 3 varying3 metrics

Grid sweep · 2×2×2. 8 runs compared across 5 hyperparameters and 3 metrics.
Run	grid · lr 1e-4 · wd 0.01 · bs 32sweep/8f3a · shard 1	grid · lr 1e-4 · wd 0.01 · bs 64sweep/8f3a · shard 2	grid · lr 1e-4 · wd 0.10 · bs 32sweep/8f3a · shard 3	grid · lr 1e-4 · wd 0.10 · bs 64sweep/8f3a · shard 4	grid · lr 3e-4 · wd 0.01 · bs 32sweep/8f3a · shard 5	grid · lr 3e-4 · wd 0.01 · bs 64sweep/8f3a · shard 6	grid · lr 3e-4 · wd 0.10 · bs 32sweep/8f3a · shard 7	grid · lr 3e-4 · wd 0.10 · bs 64sweep/8f3a · shard 8
Hyperparameters	3 of 5 differ between these runs
Learning ratevaries	0.0001	0.0001	0.0001	0.0001	0.0003	0.0003	0.0003	0.0003
Weight decayvaries	0.01	0.01	0.1	0.1	0.01	0.01	0.1	0.1
Batch size (per device)varies	32	64	32	64	32	64	32	64
Attention dropout	0.05	0.05	0.05	0.05	0.05	0.05	0.05	0.05
Random seed	7	7	7	7	7	7	7	7
Metrics	Best value marked per row; deltas are against the baseline
	88.12%	88.94%	89.03%	89.61%	90.42%best	90.18%	89.77%	87.65%
	0.451	0.428	0.421	0.409	0.372best	0.381	0.396	0.487
	90 min	70 min	91 min	71 min	89 min	69 minbest	92 min	72 min

Ready · a single run — one run is a record, not a comparison: baseline controls, varies badges, best marks, deltas and sorting are all withheld rather than measuring the run against itself.

Run detail · sft-baseline-v3

1 run10 hyperparameters5 metrics

Run detail · sft-baseline-v3. 1 run compared across 10 hyperparameters and 5 metrics.
Run	sft-baseline-v3main · 4a91c2e
Hyperparameters	10 values recorded for this run
Learning rate	0.0001
Batch size (per device)	32
Epochs	3
Optimizer	adamw
LR scheduleApplied after warmup	cosine
Warmup ratio	0.03
Adam epsilon	1e-8
Max sequence length (tokens)	4,096
Gradient checkpointingTrades compute for memory	false
Random seedNot logged before the tracker upgrade	1,337
Metrics	5 metrics recorded for this run
Validation accuracyhigher is better	89.41%
Validation losslower is better	0.412
Throughput (tokens/s)higher is better	1,840
Wall-clock training timelower is better	78 min
Peak GPU memory (GB)lower is betterThe 64-batch run crashed before the profiler flushed	38.4

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

Run comparison. 5 runs compared across 10 hyperparameters and 5 metrics, measured against baseline sft-baseline-v3.
Run	sft-baseline-v3main · 4a91c2e	lr-sweep-3e-4exp/lr-sweep · 8d1f077	lr-sweep-5e-5exp/lr-sweep · 8d1f077	long-context-8kexp/long-ctx · c72b410	batch-64-ckptexp/throughput · e05d9aa
Hyperparameters	6 of 10 differ between these runs
Learning ratevaries	0.0001	0.0003	0.00005	0.0001	0.0001
Batch size (per device)varies	32	32	32	16	64
Epochs	3	3	3	3	3
Optimizer	adamw	adamw	adamw	adamw	adamw
LR schedulevariesApplied after warmup	cosine	cosine	cosine	linear	cosine
Warmup ratio	0.03	0.03	0.03	0.03	0.03
Adam epsilon	1e-8	1e-8	1e-8	1e-8	1e-8
Max sequence length (tokens)varies	4,096	4,096	4,096	8,192	4,096
Gradient checkpointingvariesTrades compute for memory	false	false	false	true	true
Random seedvariesNot logged before the tracker upgrade	1,337	1,337	1,337	1,337	Not recorded
Metrics	Best value marked per row; deltas are against the baseline
	89.41%	91.28%+1.87 pp · +2.1%	88.02%-1.39 pp · -1.6%	91.35%best+1.94 pp · +2.2%	90.10%+0.69 pp · +0.8%
	0.412	0.361-0.051 · -12.4%	0.448+0.036 · +8.7%	0.357best-0.055 · -13.3%	0.389-0.023 · -5.6%
	1,840	1,795-45 · -2.4%	1,852+12 · +0.7%	1,120-720 · -39.1%	2,310best+470 · +25.5%
	78 min	79 min+1 min · +1.3%	77 min-1 min · -1.3%	154 min+1.3 h · +97.4%	52 minbest-26 min · -33.3%
	38.4	38.6+0.2 · +0.5%	38.2best-0.2 · -0.5%	71.9+33.5 · +87.2%	Not recorded

Run

sft-baseline-v3main · 4a91c2e

lr-sweep-3e-4exp/lr-sweep · 8d1f077

lr-sweep-5e-5exp/lr-sweep · 8d1f077

long-context-8kexp/long-ctx · c72b410

batch-64-ckptexp/throughput · e05d9aa

Hyperparameters

6 of 10 differ between these runs

Learning ratevaries

0.0001

0.0003

0.00005

0.0001

Batch size (per device)varies

Epochs

Optimizer

adamw

LR schedulevariesApplied after warmup

cosine

linear

cosine

Warmup ratio

0.03

Adam epsilon

1e-8

Max sequence length (tokens)varies

4,096

8,192

4,096

Gradient checkpointingvariesTrades compute for memory

false

true

Random seedvariesNot logged before the tracker upgrade

1,337

Not recorded

Metrics

Best value marked per row; deltas are against the baseline

89.41%

91.28%+1.87 pp · +2.1%

88.02%-1.39 pp · -1.6%

91.35%best+1.94 pp · +2.2%

90.10%+0.69 pp · +0.8%

0.412

0.361-0.051 · -12.4%

0.448+0.036 · +8.7%

0.357best-0.055 · -13.3%

0.389-0.023 · -5.6%

1,840

1,795-45 · -2.4%

1,852+12 · +0.7%

1,120-720 · -39.1%

2,310best+470 · +25.5%

78 min

79 min+1 min · +1.3%

77 min-1 min · -1.3%

154 min+1.3 h · +97.4%

52 minbest-26 min · -33.3%

38.4

38.6+0.2 · +0.5%

38.2best-0.2 · -0.5%

71.9+33.5 · +87.2%

Not recorded

Fine-tuning sweep · week 32. 5 runs compared across 10 hyperparameters and 5 metrics, measured against baseline sft-baseline-v3.
Run	sft-baseline-v3main · 4a91c2e	long-context-8kexp/long-ctx · c72b410	lr-sweep-3e-4exp/lr-sweep · 8d1f077	batch-64-ckptexp/throughput · e05d9aa	lr-sweep-5e-5exp/lr-sweep · 8d1f077
Hyperparameters	6 of 10 differ between these runs
Learning ratevaries	0.0001	0.0001	0.0003	0.0001	0.00005
Batch size (per device)varies	32	16	32	64	32
LR schedulevariesApplied after warmup	cosine	linear	cosine	cosine	cosine
Max sequence length (tokens)varies	4,096	8,192	4,096	4,096	4,096
Gradient checkpointingvariesTrades compute for memory	false	true	false	true	false
Random seedvariesNot logged before the tracker upgrade	1,337	1,337	1,337	Not recorded	1,337
Metrics	Best value marked per row; deltas are against the baseline
	89.41%	91.35%best+1.94 pp · +2.2%	91.28%+1.87 pp · +2.1%	90.10%+0.69 pp · +0.8%	88.02%-1.39 pp · -1.6%
	0.412	0.357best-0.055 · -13.3%	0.361-0.051 · -12.4%	0.389-0.023 · -5.6%	0.448+0.036 · +8.7%
	1,840	1,120-720 · -39.1%	1,795-45 · -2.4%	2,310best+470 · +25.5%	1,852+12 · +0.7%
	78 min	154 min+1.3 h · +97.4%	79 min+1 min · +1.3%	52 minbest-26 min · -33.3%	77 min-1 min · -1.3%
	38.4	71.9+33.5 · +87.2%	38.6+0.2 · +0.5%	Not recorded	38.2best-0.2 · -0.5%

Run

sft-baseline-v3main · 4a91c2e

long-context-8kexp/long-ctx · c72b410

lr-sweep-3e-4exp/lr-sweep · 8d1f077

batch-64-ckptexp/throughput · e05d9aa

lr-sweep-5e-5exp/lr-sweep · 8d1f077

Hyperparameters

6 of 10 differ between these runs

Learning ratevaries

0.0001

0.0003

0.0001

0.00005

Batch size (per device)varies

LR schedulevariesApplied after warmup

cosine

linear

cosine

Max sequence length (tokens)varies

4,096

8,192

4,096

Gradient checkpointingvariesTrades compute for memory

false

true

false

true

false

Random seedvariesNot logged before the tracker upgrade

1,337

Not recorded

1,337

Metrics

Best value marked per row; deltas are against the baseline

89.41%

91.35%best+1.94 pp · +2.2%

91.28%+1.87 pp · +2.1%

90.10%+0.69 pp · +0.8%

88.02%-1.39 pp · -1.6%

0.412

0.357best-0.055 · -13.3%

0.361-0.051 · -12.4%

0.389-0.023 · -5.6%

0.448+0.036 · +8.7%

1,840

1,120-720 · -39.1%

1,795-45 · -2.4%

2,310best+470 · +25.5%

1,852+12 · +0.7%

78 min

154 min+1.3 h · +97.4%

79 min+1 min · +1.3%

52 minbest-26 min · -33.3%

77 min-1 min · -1.3%

38.4

71.9+33.5 · +87.2%

38.6+0.2 · +0.5%

Not recorded

38.2best-0.2 · -0.5%

Grid sweep · 2×2×2. 8 runs compared across 5 hyperparameters and 3 metrics.
Run	grid · lr 1e-4 · wd 0.01 · bs 32sweep/8f3a · shard 1	grid · lr 1e-4 · wd 0.01 · bs 64sweep/8f3a · shard 2	grid · lr 1e-4 · wd 0.10 · bs 32sweep/8f3a · shard 3	grid · lr 1e-4 · wd 0.10 · bs 64sweep/8f3a · shard 4	grid · lr 3e-4 · wd 0.01 · bs 32sweep/8f3a · shard 5	grid · lr 3e-4 · wd 0.01 · bs 64sweep/8f3a · shard 6	grid · lr 3e-4 · wd 0.10 · bs 32sweep/8f3a · shard 7	grid · lr 3e-4 · wd 0.10 · bs 64sweep/8f3a · shard 8
Hyperparameters	3 of 5 differ between these runs
Learning ratevaries	0.0001	0.0001	0.0001	0.0001	0.0003	0.0003	0.0003	0.0003
Weight decayvaries	0.01	0.01	0.1	0.1	0.01	0.01	0.1	0.1
Batch size (per device)varies	32	64	32	64	32	64	32	64
Attention dropout	0.05	0.05	0.05	0.05	0.05	0.05	0.05	0.05
Random seed	7	7	7	7	7	7	7	7
Metrics	Best value marked per row; deltas are against the baseline
	88.12%	88.94%	89.03%	89.61%	90.42%best	90.18%	89.77%	87.65%
	0.451	0.428	0.421	0.409	0.372best	0.381	0.396	0.487
	90 min	70 min	91 min	71 min	89 min	69 minbest	92 min	72 min

Run

grid · lr 1e-4 · wd 0.01 · bs 32sweep/8f3a · shard 1

grid · lr 1e-4 · wd 0.01 · bs 64sweep/8f3a · shard 2

grid · lr 1e-4 · wd 0.10 · bs 32sweep/8f3a · shard 3

grid · lr 1e-4 · wd 0.10 · bs 64sweep/8f3a · shard 4

grid · lr 3e-4 · wd 0.01 · bs 32sweep/8f3a · shard 5

grid · lr 3e-4 · wd 0.01 · bs 64sweep/8f3a · shard 6

grid · lr 3e-4 · wd 0.10 · bs 32sweep/8f3a · shard 7

grid · lr 3e-4 · wd 0.10 · bs 64sweep/8f3a · shard 8

Hyperparameters

3 of 5 differ between these runs

Learning ratevaries

0.0001

0.0003

Weight decayvaries

0.01

0.1

0.01

0.1

Batch size (per device)varies

Attention dropout

0.05

Random seed

Metrics

Best value marked per row; deltas are against the baseline

88.12%

88.94%

89.03%

89.61%

90.42%best

90.18%

89.77%

87.65%

0.451

0.428

0.421

0.409

0.372best

0.381

0.396

0.487

90 min

70 min

91 min

71 min

89 min

69 minbest

92 min

72 min

Run

sft-baseline-v3main · 4a91c2e

Hyperparameters

10 values recorded for this run

Learning rate

0.0001

Batch size (per device)

Epochs

Optimizer

adamw

LR scheduleApplied after warmup

cosine

Warmup ratio

0.03

Adam epsilon

1e-8

Max sequence length (tokens)

4,096

Gradient checkpointingTrades compute for memory

false

Random seedNot logged before the tracker upgrade

1,337

Metrics

5 metrics recorded for this run

Validation accuracyhigher is better

89.41%

Validation losslower is better

0.412

Throughput (tokens/s)higher is better

1,840

Wall-clock training timelower is better

78 min

Peak GPU memory (GB)lower is betterThe 64-batch run crashed before the profiler flushed

38.4