Copy this worksheet before every training experiment. Fill values from configuration files, reports, startup JSON, epoch output, final JSON, checkpoints, and generated samples. Use <not run> or <not available> instead of leaving an ambiguous blank.
Experiment question
What specific question is this run intended to answer?
<your question>
Experiment path: <controlled comparison / larger-data training / pipeline verification / other>
What is the one main variable you intend to study?
<your variable>
Artifact identity
| Field | Value |
|---|---|
| Date and start time | <value> |
| Repository revision or workspace identifier | <value> |
| Full command | <paste command> |
| Corpus YAML | <path> |
| Corpus JSONL | <path> |
| Corpus report | <path> |
| Accepted documents | <report value> |
| Accepted characters | <report value> |
| Source contributions | <report values> |
| Exact duplicates removed | <report value> |
| Near duplicates removed | <report value> |
| Tokenizer | <path> |
| Vocabulary size | <value> |
| Encoding report | <path> |
| Training token file | <path> |
| Validation token file | <path> |
| Training documents | <encoding report value> |
| Validation documents | <encoding report value> |
| Training tokens | <encoding report value> |
| Validation tokens | <encoding report value> |
| Split seed | <encoding report value> |
Model and runtime
| Field | Value |
|---|---|
| Iteration | <1 / 2 / 3> |
| Model profile | <path> |
| Parameter count | <startup value> |
| Sequence length | <startup value> |
| Embedding dimension | <startup value> |
| Layers | <startup value> |
| Attention heads | <startup value> |
| Feed-forward dimension | <startup value> |
| CPU | <model> |
| GPU | <model or none> |
| GPU memory | <value> |
| PyTorch version | <value> |
| Device | <startup value> |
| Precision | <startup value> |
| Tensor Cores enabled | <startup value> |
| TF32 enabled | <startup value> |
| Fused AdamW | <startup value> |
| Batch size | <resolved value> |
| Workers | <resolved value> |
| Peak accelerator memory | <measured value> |
Planned training work
| Field | Value |
|---|---|
| Epoch target | <value> |
| Training example limit | <value; 0 means all> |
| Validation example limit | <value; 0 means all> |
| Resolved training examples | <startup value> |
| Resolved validation examples | <startup value> |
| Training batches per epoch | <startup value> |
| Maximum training steps | <startup value> |
| Learning rate | <value> |
| Weight decay | <value> |
| Estimated steps per second | <bounded measurement> |
| Estimated tokens per second | <bounded measurement> |
| Estimated duration | <calculation and allowance> |
| Checkpoint directory | <path> |
| Free storage before run | <value> |
| Backup destination | <path or policy> |
Epoch record
| Epoch | Training loss | Validation loss | Ending learning rate | Global step | Processed tokens | Throughput | Elapsed time | Checkpoint path | Notes |
|---|---|---|---|---|---|---|---|---|---|
| 1 | <value> | <value> | <value> | <value> | <value> | <value> | <value> | <path> | <observation> |
| 2 | <value> | <value> | <value> | <value> | <value> | <value> | <value> | <path> | <observation> |
| 3 | <value> | <value> | <value> | <value> | <value> | <value> | <value> | <path> | <observation> |
Add rows for additional epochs.
Completion or interruption
| Final JSON field | Value |
|---|---|
training_complete | <value> |
interrupted | <value> |
completed_epochs | <value> |
global_step | <value> |
best_validation_loss | <value> |
verified_checkpoint | <path> |
restored_epoch | <value> |
restored_checkpoint_type | <value> |
If resumed:
| Resume field | Value |
|---|---|
resumed_from | <path> |
restored_epoch | <value> |
restored_global_step | <value> |
checkpoint_type | <value> |
| Settings compared with original | <confirmation or difference> |
Fixed generation protocol
| Setting | Value |
|---|---|
| Checkpoint | <path> |
| Tokenizer | <path> |
| Device | <value> |
| Maximum new tokens | <value> |
| Greedy | <true / false> |
| Temperature | <value> |
| Top-k | <value> |
| Top-p | <value> |
| Repetition penalty | <value> |
| No-repeat n-gram size | <value> |
| Seed | <value> |
| Stop settings | <value> |
Generated evidence
| Prompt | Full unedited output location | Finish reason | Generated tokens | Relevance | Grammar | Repetition | Factual issues |
|---|---|---|---|---|---|---|---|
| The purpose of a firewall is | <paste or link> | <value> | <value> | <observation> | <observation> | <observation> | <observation> |
| When a software service fails in production, | <paste or link> | <value> | <value> | <observation> | <observation> | <observation> | <observation> |
| In the history of computing, | <paste or link> | <value> | <value> | <observation> | <observation> | <observation> | <observation> |
| Water changes from a liquid to | <paste or link> | <value> | <value> | <observation> | <observation> | <observation> | <observation> |
| The evidence does not prove that | <paste or link> | <value> | <value> | <observation> | <observation> | <observation> | <observation> |
Conclusion
Did the engineering pipeline work?
<evidence-based answer>
What did training and validation show?
<evidence-based answer>
What did generation show?
<evidence-based answer>
What can this experiment support as a conclusion?
<specific conclusion>
What can this experiment not support?
<limitation>
What is the next experiment, and which evidence justifies it?
<one-variable experiment and reason>
Return to the course home or the model comparison lesson.