minicpm5-finetune-ms-swift

Fine-tune MiniCPM5-1B or MiniCPM5-2B with ms-swift or Megatron-SWIFT. Use when the user mentions "ms-swift", "swift sft", "swift rlhf", or "megatron sft". MiniCPM5-1B uses the PyPI release with `--template minicpm5`; MiniCPM5-2B requires `ms-swift>=4.6.0.dev0` with `--template minicpm5_2b`. Both use `--model_type llama`.

Quality

84%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Passed

No findings from the security scan

Quality

Content

75%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 well-organized and highly actionable, with concrete commands throughout and a real validation step. It sits just below top marks due to inline version pins, ellipsis-truncated secondary command variants, a passive validation step without error-recovery guidance, and no bundle reference files.

Suggestions

Add an error-recovery feedback loop to the Validate step (e.g., 'if loss does not decrease after N steps, lower learning_rate or check --template / dataset format').

Replace the ellipsis-truncated Full SFT/DPO/Megatron snippets with complete executable commands, or explicitly justify the abbreviation.

Move pinned version numbers and the commit hash into a dedicated 'Version pins / reproducibility' or 'deprecated' section so time-sensitive details don't burden the active install steps.

Dimension	Reasoning	Score
Conciseness	The body is lean — a version table, install/train/merge commands, a short validate snippet, and a pitfalls list — with no padding of concepts Claude already knows. Not 5 because time-sensitive version pins ('ms-swift==4.5.3', 'transformers>=5.6,<5.17', the commit hash, 'mcore-bridge==1.6.4') sit inline in the active steps rather than in a deprecated/old-patterns section, which the rubric flags as a conciseness penalty.	4 / 5
Actionability	The primary LoRA SFT command and the install/merge commands are fully executable and copy-paste ready with all flags. Not 5 because the Full SFT/DPO/Megatron variants use ellipsis truncation ('swift sft --tuner_type full ...', '...') rather than complete commands, leaving minor gaps in the secondary paths.	4 / 5
Workflow Clarity	Steps are clearly numbered (1. Install, 2. Train, 3. Validate) with a validation checkpoint showing expected loss decay. Not 5 because the validate step is passive ('Loss should decrease') with no error-recovery feedback loop (no 'if loss plateaus, do X'). The batch-operation cap-at-3 does not apply since an explicit validation step is present.	4 / 5
Progressive Disclosure	Content is well-sectioned (Required input, Steps, Merge, Full SFT/DPO/RLHF, Multi-GPU, Common pitfalls, Reference) with a single clearly-signaled one-level reference to docs/finetune/ms_swift.md plus external doc links; no nested references. Not 5 because no bundle reference files exist to offload the detail, so most content is inline rather than split across a reference layer.	4 / 5
	Total	16 / 20 Passed

Description

87%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 is strong: third-person voice, explicit what-and-when with concrete trigger terms, and a clearly delineated niche. It loses only minor points for not enumerating a broader set of action verbs and synonyms.

Dimension	Reasoning	Score
Specificity	Quotes 'Fine-tune MiniCPM5-1B or MiniCPM5-2B with ms-swift or Megatron-SWIFT' plus concrete template/version configs ('--template minicpm5', '--template minicpm5_2b', '--model_type llama') list several specific actions across two models and frameworks, with minor gaps (e.g. full SFT/DPO/RLHF variants are implied by 'fine-tune' rather than enumerated). Not 5 because the core action verb is essentially one ('fine-tune') rather than multiple distinct concrete actions.	4 / 5
Completeness	Clearly answers both: what ('Fine-tune MiniCPM5-1B or MiniCPM5-2B with ms-swift or Megatron-SWIFT') and when ('Use when the user mentions "ms-swift", "swift sft", "swift rlhf", or "megatron sft"') with concrete trigger phrases, matching the anchor exactly.	5 / 5
Trigger Term Quality	Quotes the trigger clause 'Use when the user mentions "ms-swift", "swift sft", "swift rlhf", or "megatron sft"' — these are exactly the natural phrases a user would say. Not 5 because natural variations like 'swift export', 'swift dpo', 'LoRA fine-tune', or 'MiniCPM5 training' are not included as synonyms.	4 / 5
Distinctiveness Conflict Risk	The niche is narrow (MiniCPM5 fine-tuning via ms-swift/Megatron-SWIFT) with distinct, product-specific triggers; minimal overlap risk with other skills. Matches the anchor for a clear niche with distinct triggers.	5 / 5
	Total	18 / 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
relative_links	Relative link issues: 1 suspicious	Warning

	Total	15 / 16 Passed

Repository: OpenBMB/MiniCPM
Path: skills/minicpm5-finetune-ms-swift/SKILL.md
Commit: 316cfb1

Reviewed: 15 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.