CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

implementing-llms-litgpt

Implements and trains LLMs using Lightning AI's LitGPT with 20+ pretrained architectures (Llama, Gemma, Phi, Qwen, Mistral). Use when need clean model implementations, educational understanding of architectures, or production fine-tuning with LoRA/QLoRA. Single-file implementations, no abstraction layers.

Quality

73%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Passed

No findings from the security scan

Fix and improve this skill with Tessl

tessl review fix ./skills/model-architecture/litgpt/SKILL.md
SKILL.md
Quality
Evals
Security

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.

A well-structured, highly actionable skill body with four clearly sequenced workflows and properly signaled, existing reference files. Main gaps are duplicated LoRA command blocks, inline comments that break bash line continuations, missing validation checkpoints in the training workflows, and references to scripts that are not part of the bundle.

Suggestions

Consolidate the three near-identical `litgpt finetune_lora` command blocks (Workflow 1 Step 3 and Workflow 2 Steps 2-3) into a single canonical example with the parameter guide, and cut the redundant LoRA rank list.

Remove inline comments placed after `\` line-continuations in bash blocks (they make the commands non-executable if copied) and move those explanations to bullet lists below each command.

Add validation checkpoints to the workflows: verify the dataset loads and matches the expected JSON schema before launching training, check training loss before merging LoRA weights, and test the merged/converted model's output before the deployment step.

Either add the referenced `scripts/prepare_dataset.py` and `scripts/convert_lit_checkpoint.py` to the bundle or replace those invocations with the equivalent built-in `litgpt` commands.

DimensionReasoningScore

Conciseness

The body is mostly commands with little concept over-explanation, but the full `litgpt finetune_lora` invocation appears three times (Workflow 1 Step 3, Workflow 2 Steps 2 and 3), the LoRA rank guide duplicates Step 2 parameter commentary, and the repeated copyable checklists add padding. Not 4 because the duplication is clearly trimmable; not 2 because there is no padded conceptual tutorial content.

3 / 5

Actionability

Nearly everything is concrete, copy-paste-ready commands or code (`litgpt download`, `finetune`, `finetune_lora`, `pretrain`, `merge_lora`, the FastAPI deployment app). Not 5 because the LoRA blocks place inline comments after `\` line continuations (which breaks execution if copied verbatim) and Workflow 3/4 reference `scripts/prepare_dataset.py` and `scripts/convert_lit_checkpoint.py` that are not in the bundle; not 3 because the guidance is overwhelmingly executable rather than pseudocode.

4 / 5

Workflow Clarity

All four workflows have clear checklists and numbered, well-sequenced steps, but validation checkpoints are missing or implicit: no dataset-format verification before training, no convergence/loss check before finishing, and no test of the merged model before deployment (only Workflow 4 Step 1 tests inference). Not 4 because these long-running batch training workflows lack explicit validate-before-proceed steps; not 2 because sequencing is coherent and the Common issues section provides some error-recovery guidance.

3 / 5

Progressive Disclosure

The Advanced topics section clearly signals four one-level-deep references (supported-models.md, training-recipes.md, distributed-training.md, custom-models.md), all of which exist in the bundle, and the core workflows are appropriately kept in the body. Not 5 because the ~460-line body inlines content (common-issues troubleshooting, hardware tables, full command variants) that could be split out; not 3 because structure and reference signaling are good.

4 / 5

Total

14

/

20

Passed

Description

75%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.

A solid description with concrete actions, a named tool, specific architectures, and an explicit Use-when clause covering the main fine-tuning use cases. The main weaknesses are the ungrammatical trigger clause and missing trigger coverage for pretraining and deployment workflows that the body documents.

Suggestions

Fix the trigger clause grammar ("Use when need" → "Use when the user needs") and extend it to cover the body's other core workflows, e.g. "pretraining models from scratch" and "converting/quantizing models for deployment".

Add natural trigger synonyms such as "pretrain", "train from scratch", "LoRA adapters", and "quantize" so users phrasing requests differently still match.

Consider dropping the trailing "Single-file implementations, no abstraction layers" fragment or folding it into the what-clause; as a standalone fragment it reads as a tagline rather than a capability.

DimensionReasoningScore

Specificity

"Implements and trains LLMs", "production fine-tuning with LoRA/QLoRA", and "20+ pretrained architectures (Llama, Gemma, Phi, Qwen, Mistral)" name several concrete actions and specific models. It is not 5 because the action list is not comprehensive (pretraining from scratch and deployment/conversion workflows from the body are absent), and not 3 because it goes well beyond 1-2 actions.

4 / 5

Completeness

Both parts are present: a clear what ("Implements and trains LLMs using Lightning AI's LitGPT with 20+ pretrained architectures") and an explicit when ("Use when need clean model implementations, educational understanding of architectures, or production fine-tuning with LoRA/QLoRA"). Not 5 because the when-clause is ungrammatical ("Use when need") and omits pretraining/deployment triggers that the body treats as core workflows; not 3 because both what and when are explicit.

4 / 5

Trigger Term Quality

Strong natural terms users would say: "LitGPT", "Llama", "Gemma", "Phi", "Qwen", "Mistral", "LoRA", "QLoRA", "fine-tuning". Not 5 because common variations like "pretrain", "train from scratch", "quantize", or "GGUF" are missing; not 3 because coverage of natural model and technique names is clearly good.

4 / 5

Distinctiveness Conflict Risk

Naming LitGPT, Lightning AI, and specific architectures creates a clear niche with distinct triggers. Not 5 because the broad framing "trains LLMs" overlaps with other fine-tuning skills (Axolotl/TRL are acknowledged in the body); not 3 because the tool-specific naming makes confusion unlikely.

4 / 5

Total

16

/

20

Passed

Validation

87%

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

Validation — 14 / 16 Passed

Validation for skill structure

CriteriaDescriptionResult

frontmatter_unknown_keys

Unknown frontmatter key(s) found; consider removing or moving to metadata

Warning

referenced_paths_exist

Referenced path issues: 2 possibly missing (scripts/convert_lit_checkpoint.py, scripts/prepare_dataset.py)

Warning

Total

14

/

16

Passed

Repository

OpenLAIR/dr-claw

Path

skills/model-architecture/litgpt/SKILL.md

Reviewed

Table of Contents

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.