Publish versioned Markdown. Build focused streams. Give people and agents only the context they need.
Full detail backing the summary in SKILL.md. Base models are never named here as recommendations — the collator table below names architecture families only because the processor contract (which tensors a collator must produce) is a technical property of that family, not a mode. Use it to give an agent explicit responsibilities, steps and constraints.
This skill assumes finetuning-method-selection already routed here: the data shape is image+text demonstrations, not preference pairs or a verifiable reward signal, and the base is a vision-language model rather than a text-only one. lora-qlora-recipes covers the text-only Lo. Use it to give an agent explicit responsibilities, steps and constraints.
Concrete JSONL-to-JSONL conversions for every pattern in SKILL.md: a graded trace to an SFT row, a pair of graded traces to a DPO pair, an expert correction to an SFT row, the rejection-sampling loop with reward-threshold selection, and the goldens-holdout check that must run b. Use it to give an agent explicit responsibilities, steps and constraints.
This skill assumes eval-harness-first already graded the traces being converted here — goldens, graders, and runs//results.json all exist before conversion starts. This is the flywheel edge that skill names in its own flow: "the same labeled traces become the training set." C. Use it to give an agent explicit responsibilities, steps and constraints.
Complete command sequences for every format on the SKILL.md Format Map, plus the smoke-test script skeleton. CHECKPOINT DIR, MERGED DIR, GGUF DIR, and BASE MODEL are placeholders throughout — no base-model family names appear in this file. Fill each with the promoted ch. Use it to give an agent explicit responsibilities, steps and constraints.
The last stop after checkpoint-promotion hands off a PROMOTE verdict: a checkpoint that cleared the four-stage gate still isn't deployed until it's exported in the right format for its target runtime and proven to still work post-export. A REJECT verdict never reaches this. Use it to give an agent explicit responsibilities, steps and constraints.
This skill assumes finetuning-method-selection already routed here because the data shape is preference pairs or unpaired thumbs-up/down feedback, not demonstrations (that's lora-qlora-recipes) or a verifiable reward signal (that's grpo-rlvr-training). What follows is metho. Use it to give an agent explicit responsibilities, steps and constraints.
Unsloth is a fast-kernel wrapper over PEFT and TRL, not a replacement API — every Unsloth kwarg below has a plain TRL/PEFT equivalent. Use this table to translate an Unsloth config to plain TRL (or back), and to know which knob lives on which object in the current TRL API. Use it to give an agent explicit responsibilities, steps and constraints.
Full tables and a complete worked config backing the summary in SKILL.md. Base models are never named here — every example is labeled by size class only; see finetuning-method-selection's references/model-catalog.md for which actual model to use at a given size class. Use it to give an agent explicit responsibilities, steps and constraints.
This skill assumes the routing decision already happened — finetuning-method-selection should have already pointed here because the data shape is demonstrations (SFT), not preference pairs or a verifiable reward signal. What follows is the current best-practice recipe for confi. Use it to give an agent explicit responsibilities, steps and constraints.
Complete, runnable reward functions for TRL's GRPOTrainer. Every function here follows the current TRL reward-function signature: it accepts completions plus any extra dataset columns as keyword arguments, and returns a list[float] the same length as completions. Base mod. Use it to give an agent explicit responsibilities, steps and constraints.
This skill assumes finetuning-method-selection already routed here because the target behavior has a verifiable pass/fail signal — not demonstrations (lora-qlora-recipes) or preference pairs (preference-optimization). What follows is when RL is the right tool, the reference. Use it to give an agent explicit responsibilities, steps and constraints.
Last verified: 2026-07-13 — refresh when a new size-class anchor is validated or optimizer/dtype defaults change. Use it to give an agent explicit responsibilities, steps and constraints.
This is the router skill for the fine-tuning lifecycle: it decides whether fine-tuning is the right tool at all, and if so, which method and which base-model size class. Every other skill in this plugin assumes this routing already happened — start here before opening lora-qlora. Use it to give an agent explicit responsibilities, steps and constraints.
The full procedure behind SKILL.md's "Judge Calibration Is a Prerequisite" section. Any grader routed to an LLM-judge follows this before its verdicts count toward a pass rate or a checkpoint promotion decision. Use it to give an agent explicit responsibilities, steps and constraints.
Runnable examples for the four grader shapes named in SKILL.md's Graders section: schema-compliance, exact-match with normalization, execution-based, and LLM-judge. Every grader returns a binary pass/fail — never a Likert score — per the plugin-wide rule. Wire each one to exact. Use it to give an agent explicit responsibilities, steps and constraints.
The Phase 0 gate for the whole plugin: finetuning-method-selection and every downstream skill assume this harness exists before a training config gets written. The harness is not a run-end side artifact — it is the data-curation engine. The same labeled traces that build the go. Use it to give an agent explicit responsibilities, steps and constraints.
Full detail backing SKILL.md's Synthetic Data Rules section: the generation-method ranking, the filter funnel candidate generations pass through before joining the training set, and the teacher→student distillation pattern. Base models are never named as recommendations here. Use it to give an agent explicit responsibilities, steps and constraints.
Concrete JSONL examples for every format in SKILL.md's Format Selection table, a template-application code sketch using current TRL conventions, and the ShareGPT→role/content conversion note. Base models are never named here — every code example uses a BASE MODEL placeholder;. Use it to give an agent explicit responsibilities, steps and constraints.
This skill assumes finetuning-method-selection already routed here — the next step is preparing data, not choosing a method. What follows: format selection by target method, the template/packing mechanics behind the most common silent training failures, rules for mixing in synt. Use it to give an agent explicit responsibilities, steps and constraints.
Publish versioned Markdown. Build focused streams. Give people and agents only the context they need.
Full detail backing the summary in SKILL.md. Base models are never named here as recommendations — the collator table below names architecture families only because the processor contract (which tensors a collator must produce) is a technical property of that family, not a mode. Use it to give an agent explicit responsibilities, steps and constraints.
This skill assumes finetuning-method-selection already routed here: the data shape is image+text demonstrations, not preference pairs or a verifiable reward signal, and the base is a vision-language model rather than a text-only one. lora-qlora-recipes covers the text-only Lo. Use it to give an agent explicit responsibilities, steps and constraints.
Concrete JSONL-to-JSONL conversions for every pattern in SKILL.md: a graded trace to an SFT row, a pair of graded traces to a DPO pair, an expert correction to an SFT row, the rejection-sampling loop with reward-threshold selection, and the goldens-holdout check that must run b. Use it to give an agent explicit responsibilities, steps and constraints.
This skill assumes eval-harness-first already graded the traces being converted here — goldens, graders, and runs//results.json all exist before conversion starts. This is the flywheel edge that skill names in its own flow: "the same labeled traces become the training set." C. Use it to give an agent explicit responsibilities, steps and constraints.
Complete command sequences for every format on the SKILL.md Format Map, plus the smoke-test script skeleton. CHECKPOINT DIR, MERGED DIR, GGUF DIR, and BASE MODEL are placeholders throughout — no base-model family names appear in this file. Fill each with the promoted ch. Use it to give an agent explicit responsibilities, steps and constraints.
The last stop after checkpoint-promotion hands off a PROMOTE verdict: a checkpoint that cleared the four-stage gate still isn't deployed until it's exported in the right format for its target runtime and proven to still work post-export. A REJECT verdict never reaches this. Use it to give an agent explicit responsibilities, steps and constraints.
This skill assumes finetuning-method-selection already routed here because the data shape is preference pairs or unpaired thumbs-up/down feedback, not demonstrations (that's lora-qlora-recipes) or a verifiable reward signal (that's grpo-rlvr-training). What follows is metho. Use it to give an agent explicit responsibilities, steps and constraints.
Unsloth is a fast-kernel wrapper over PEFT and TRL, not a replacement API — every Unsloth kwarg below has a plain TRL/PEFT equivalent. Use this table to translate an Unsloth config to plain TRL (or back), and to know which knob lives on which object in the current TRL API. Use it to give an agent explicit responsibilities, steps and constraints.
Full tables and a complete worked config backing the summary in SKILL.md. Base models are never named here — every example is labeled by size class only; see finetuning-method-selection's references/model-catalog.md for which actual model to use at a given size class. Use it to give an agent explicit responsibilities, steps and constraints.
This skill assumes the routing decision already happened — finetuning-method-selection should have already pointed here because the data shape is demonstrations (SFT), not preference pairs or a verifiable reward signal. What follows is the current best-practice recipe for confi. Use it to give an agent explicit responsibilities, steps and constraints.
Complete, runnable reward functions for TRL's GRPOTrainer. Every function here follows the current TRL reward-function signature: it accepts completions plus any extra dataset columns as keyword arguments, and returns a list[float] the same length as completions. Base mod. Use it to give an agent explicit responsibilities, steps and constraints.
This skill assumes finetuning-method-selection already routed here because the target behavior has a verifiable pass/fail signal — not demonstrations (lora-qlora-recipes) or preference pairs (preference-optimization). What follows is when RL is the right tool, the reference. Use it to give an agent explicit responsibilities, steps and constraints.
Last verified: 2026-07-13 — refresh when a new size-class anchor is validated or optimizer/dtype defaults change. Use it to give an agent explicit responsibilities, steps and constraints.
This is the router skill for the fine-tuning lifecycle: it decides whether fine-tuning is the right tool at all, and if so, which method and which base-model size class. Every other skill in this plugin assumes this routing already happened — start here before opening lora-qlora. Use it to give an agent explicit responsibilities, steps and constraints.
The full procedure behind SKILL.md's "Judge Calibration Is a Prerequisite" section. Any grader routed to an LLM-judge follows this before its verdicts count toward a pass rate or a checkpoint promotion decision. Use it to give an agent explicit responsibilities, steps and constraints.
Runnable examples for the four grader shapes named in SKILL.md's Graders section: schema-compliance, exact-match with normalization, execution-based, and LLM-judge. Every grader returns a binary pass/fail — never a Likert score — per the plugin-wide rule. Wire each one to exact. Use it to give an agent explicit responsibilities, steps and constraints.
The Phase 0 gate for the whole plugin: finetuning-method-selection and every downstream skill assume this harness exists before a training config gets written. The harness is not a run-end side artifact — it is the data-curation engine. The same labeled traces that build the go. Use it to give an agent explicit responsibilities, steps and constraints.
Full detail backing SKILL.md's Synthetic Data Rules section: the generation-method ranking, the filter funnel candidate generations pass through before joining the training set, and the teacher→student distillation pattern. Base models are never named as recommendations here. Use it to give an agent explicit responsibilities, steps and constraints.
Concrete JSONL examples for every format in SKILL.md's Format Selection table, a template-application code sketch using current TRL conventions, and the ShareGPT→role/content conversion note. Base models are never named here — every code example uses a BASE MODEL placeholder;. Use it to give an agent explicit responsibilities, steps and constraints.
This skill assumes finetuning-method-selection already routed here — the next step is preparing data, not choosing a method. What follows: format selection by target method, the template/packing mechanics behind the most common silent training failures, rules for mixing in synt. Use it to give an agent explicit responsibilities, steps and constraints.