Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Minimal GPT pretraining and fine-tuning (nanoGPT). The simplest, fastest repository for training medium-sized GPTs with ~300-line model.py and ~300-line train.py. Reproduces GPT-2 (124M) on OpenWebText. Supports DDP multi-GPU/multi-node, character-level training, weight loading from HuggingFace GPT-2 checkpoints, and simple finetuning. Note: superseded by nanochat for new projects; this repo remains valuable as a reference implementation and learning tool.
.claude/skills/mkurman-nanogpt/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-16 | ✗→✓ | ▲ Improved | 153% | 0% |
| case-20 | ✗→✓ | ▲ Improved | -62% | 0% |
| case-01 | ✓→✓ | ＝ Same ✓ | 66% | 0% |
| case-02 | ✓→✓ | ＝ Same ✓ | 79% | 0% |
| case-03 | ✓→✓ | ＝ Same ✓ | 34% | 0% |
--|--------|--------|-------|---------|-------------| | Shakespeare char | ~10M | 6 | 6 | 384 | 1 GPU, 3 min | | GPT-2 small | 124M | 12 | 12 | 768 | 8×A100, 4 days | | GPT-2 medium | 350M | 24 | 16 | 1024 | Modify config | | GPT-2 large | 774M | 36 | 20 | 1280 | Modify config | | GPT-2 XL | 1.5B | 48 | 25 | 1600 | Multi-node |
bash# Benchmark training speed python train.py config/train_gpt2.py --eval_only # Loss tracking # nanoGPT-reported GPT-2 (124M) on OpenWebText: val loss 2.85 # OpenAI GPT-2 (124M) on WebText: val loss ~3.11 # (domain gap between WebText and OpenWebText accounts for difference) # Custom benchmarks python train.py config/train_shakespeare_char.py --eval_iters=200
init_from='scratch'|'resume'|'gpt2*' — switch between training modes--compile=True on GPU for ~2x speedupgradient_accumulation_steps emulates larger batch sizeswte (embedding) and lm_head — standard GPT practicetorchrun for multi-GPU — not python -m torch.distributed.launch| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | pass→pass | 4,302 | 3,268 | -24% | 1 | 1 | 0% | 752 | 1,249 | +66% | 0 | 0 | — |
case-02 | pass→pass | 4,069 | 3,720 | -9% | 1 | 1 | 0% | 716 | 1,282 | +79% | 0 | 0 | — |
case-03 | pass→pass | 5,540 | 3,596 | -35% | 1 | 1 | 0% | 922 | 1,239 | +34% | 0 | 0 | — |
case-04 | pass→pass | 4,152 | 2,562 | -38% | 1 | 1 | 0% | 694 | 1,130 | +63% | 0 | 0 | — |
case-05 | pass→pass | 5,534 | 2,581 | -53% | 1 | 1 | 0% | 965 | 1,113 | +15% | 0 | 0 | — |
case-06 | pass→pass | 12,593 | 9,731 | -23% | 1 | 1 | 0% | 2,214 | 2,390 | +8% | 0 | 0 | — |
case-07 | pass→pass | 5,421 | 3,739 | -31% | 1 | 1 | 0% | 999 | 1,343 | +34% | 0 | 0 | — |
case-08 | pass→pass | 3,287 | 2,567 | -22% | 1 | 1 | 0% | 548 | 1,086 | +98% | 0 | 0 | — |
case-09 | pass→pass | 2,919 | 2,866 | -2% | 1 | 1 | 0% | 472 | 1,106 | +134% | 0 | 0 | — |
case-10 | pass→pass | 5,398 | 4,186 | -22% | 1 | 1 | 0% | 1,037 | 1,411 | +36% | 0 | 0 | — |
case-11 | pass→pass | 9,916 | 4,529 | -54% | 1 | 1 | 0% | 1,588 | 1,362 | -14% | 0 | 0 | — |
case-12 | pass→pass | 10,069 | 10,949 | +9% | 1 | 1 | 0% | 1,812 | 2,651 | +46% | 0 | 0 | — |
case-13 | pass→pass | 5,191 | 3,137 | -40% | 1 | 1 | 0% | 806 | 1,198 | +49% | 0 | 0 | — |
case-14 | pass→pass | 4,326 | 2,682 | -38% | 1 | 1 | 0% | 721 | 1,053 | +46% | 0 | 0 | — |
case-15 | pass→pass | 3,032 | 2,545 | -16% | 1 | 1 | 0% | 451 | 969 | +115% | 0 | 0 | — |
case-16 | fail→pass | 5,671 | 8,081 | +42% | 1 | 1 | 0% | 822 | 2,080 | +153% | 0 | 0 | — |
case-17 | pass→pass | 14,201 | 16,012 | +13% | 1 | 1 | 0% | 2,263 | 3,121 | +38% | 0 | 0 | — |
case-18 | pass→pass | 6,962 | 3,667 | -47% | 1 | 1 | 0% | 1,314 | 1,308 | -0% | 0 | 0 | — |
case-19 | pass→pass | 5,064 | 2,929 | -42% | 1 | 1 | 0% | 814 | 1,163 | +43% | 0 | 0 | — |
case-20 | fail→pass | 15,422 | 2,663 | -83% | 1 | 1 | 0% | 2,840 | 1,065 | -63% | 0 | 0 | — |
case-21 | pass→pass | 13,819 | 15,433 | +12% | 1 | 1 | 0% | 2,645 | 3,584 | +36% | 0 | 0 | — |
case-22 | pass→pass | 13,775 | 16,433 | +19% | 1 | 1 | 0% | 2,551 | 3,743 | +47% | 0 | 0 | — |
case-23 | pass→pass | 18,882 | 22,516 | +19% | 1 | 1 | 0% | 3,064 | 4,238 | +38% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 23 cases were attempted. The headline lift of +9 percentage points is the difference between those two pass rates over the 23 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.