nanogpt

Educational GPT implementation in ~300 lines. Reproduces GPT-2 (124M) on OpenWebText. Clean, hackable code for learning transformers. By Andrej Karpathy. Perfect for understanding GPT architecture from scratch. Train on Shakespeare (CPU) or OpenWebText (multi-GPU).

Quality

57%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Low

Low-risk findings worth noting

Fix and improve this skill with Tessl

tessl review fix ./backend/cli/skills/ml-training/nanogpt/SKILL.md

Quality

Content

61%Weight 40%Scale 1-5

Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.

The body is a well-structured, actionable guide with real progressive-disclosure references, but it lacks validation/verification checkpoints in its training workflows and carries some over-explanation of familiar concepts. Strong on actionability, moderate on conciseness and workflow clarity.

Suggestions

Add validation checkpoints to workflows (e.g. 'After prepare.py, verify train.bin/val.bin exist and are non-empty'; 'After training, run sample.py to confirm generation quality before scaling up') to lift workflow clarity above the batch-operation cap.

Trim familiar restatements (config comments like '# 6 transformer layers', the 'Simplicity advantages' bullet list, and obvious hardware notes) to improve conciseness.

Complete the custom-dataset example by including the meta.pkl/encoder save step so the prepare.py snippet is fully executable end-to-end.

Dimension	Reasoning	Score
Conciseness	Mostly efficient with executable snippets, but it re-explains familiar concepts ('Hackable', 'No abstractions', config comments restating obvious values) and inlines full config blocks that could be trimmed; some padding present.	3 / 5
Actionability	Provides concrete, mostly copy-paste-ready commands and configs across Shakespeare, GPT-2, fine-tuning, and custom datasets, with only minor gaps (e.g. the custom-dataset prepare.py lacks the meta.pkl save step).	4 / 5
Workflow Clarity	Multi-step workflows are clearly sequenced (prepare -> train -> sample) but they involve batch/multi-GPU training with no validation or verification checkpoints, which caps workflow clarity at 3 per the destructive/batch guidance.	3 / 5
Progressive Disclosure	SKILL.md is a clear overview with well-signaled one-level-deep references to real files (architecture.md, training.md, data.md) that exist in references/; content is appropriately split with only minor inlining of config detail.	4 / 5
	Total	14 / 20 Passed

Description

53%Weight 40%Scale 1-5

Based on the skill's description, can an agent find and select it at the right time? Clear, specific descriptions lead to better discovery.

The description concretely identifies the skill and its domain but reads as a feature blurb rather than trigger-oriented guidance, omitting any explicit 'Use when...' clause. It is specific enough to be distinct but lacks the natural trigger phrases and completeness of high-quality examples.

Suggestions

Add an explicit trigger clause such as 'Use when learning GPT/transformer architecture, reproducing GPT-2, or training small character-level models'.

Surface natural user phrasings (e.g. 'train a GPT from scratch', 'fine-tune GPT-2', 'learn transformers') to improve trigger-term quality.

Reframe descriptive attributes ('Perfect for understanding...') as concrete actions to lift specificity from naming the domain to listing actionable capabilities.

Dimension	Reasoning	Score
Specificity	Names the domain and concrete actions (reproduce GPT-2, train on Shakespeare/OpenWebText) but is more descriptive of what the code is than a list of executable actions, leaving coverage partial.	3 / 5
Completeness	It clearly states what the skill is (educational GPT implementation) but lacks any 'Use when...' trigger clause, so 'when' is only weakly implied; per guidelines a missing explicit trigger caps completeness at 3.	3 / 5
Trigger Term Quality	Includes relevant terms (GPT, GPT-2, transformer, training, Shakespeare, multi-GPU) but lacks natural user phrasings like 'train a GPT', 'fine-tune', 'learn transformers' and omits common synonyms.	3 / 5
Distinctiveness Conflict Risk	The GPT-2/nanoGPT/niche with named author and datasets makes it mostly distinct with only minor overlap risk against general transformer training skills.	4 / 5
	Total	13 / 20 Passed

Validation

93%

Warnings & errors only

Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.

Validation — 15 / 16 Passed

Validation for skill structure

Criteria	Description	Result
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning

	Total	15 / 16 Passed

Repository: synthetic-sciences/openscience
Path: backend/cli/skills/ml-training/nanogpt/SKILL.md
Commit: fcd2e11

Reviewed: 2 days ago

Table of Contents

Discovery Implementation Validation

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.