Training loss
Actual training objective
Training intelligence, one rule at a time
A clearer view of learning.
Follow Qwen as it learns to write programs and predict ARC grids.
Optimizer steps
—
Waiting for telemetry
Training loss
—
Completion-token NLL
Validation loss
—
Held-out synthetic episodes
Next loss check
—
Every 100 optimizer steps
Learning curves
Separate scales. Actual recorded values. No smoothing.
EVERY STEP
Validation loss
Weighted validation loss on 128 fixed synthetic episodes
Program · 1×
Reference program token mean · actual weighted contribution
Format · 5×
JSON brackets, separators, fields and code fences · weighted contribution
Shape · 5×
Actual weighted contribution
Grid cells · 5×
Actual weighted contribution
Program + format×5 + cells×5. No separate size loss. Component charts show weighted contributions, on separate scales.
01 / Quick loss check
Next at step 100Baseline at step 0, then every 100 steps.
02 / Checkpoint + Hub upload
Next at step 1,000Every 1,000 steps; final adapter at completion.
03 / Generated solutions
Steps 1,000 · 2,000 · 3,000 · finalThinking on · all 20,471 available validation episodes. Fresh weighted SFT run; previous-run scores are archived.
Learning rate
100 warmup steps, followed by cosine decay
Run configuration
- Training episodes
- —
- Task families
- —
- Global batch
- —
- Planned epochs
- 1
- Latest saved checkpoint
- None yet
- Training mode
- Weighted program + grid SFT
Loss measures prediction of saved target tokens. Solving accuracy requires generating a fresh answer and executing its program.
Family progress
Which rules are getting learned? Inspect each family across checkpoints.
Loading family inventory…
↕ Scroll rows · ↔ Scroll columns · drag either scrollbar
| Family | Train / val episodes | Evaluated | Program exact | Grid exact | Both exact | Token limit | Change |
|---|
—
Select a family to inspect its size distribution and checkpoint scores.
Step by step
The latest 12 measurements, in chronological order.
↕ Scroll through step metrics · ↔ Scroll for additional columns
| Step | Total loss | Program | Format ×5 | Size ×5 | Cells ×5 | Val loss | Step seconds | Learning rate |
|---|---|---|---|---|---|---|---|---|
| Waiting for recorded measurements… | ||||||||
Hardware telemetry · 8 × H200
Sampled every two minutes. Utilization is a hardware measurement, not evidence of an optimizer update.