asc-ml-adaptation

Files (2)

• SKILL.md
• openai.yaml · agents/

SKILL.md (first lines)

# Asc Ml Adaptation

Pick the smallest adaptation that can plausibly move the model.
Do not jump to the heaviest training method first.

## Decision Order

### 1. CPT

Use continued pretraining when the model needs better domain or language coverage before task behavior shaping.

Use CPT for:
- in-domain terminology
- corpus style and distribution
- broad language coverage
- domain recall before instruction tuning

Look for:
- held-out loss or perplexity
- terminology accuracy
- downstream sanity checks
- slice regressions in the target domain

### 2. SFT

Use supervised fine-tuning when the model needs to learn the task shape directly.

Use SFT for:
- response structure
- policy-following behavior
- formatting conventions
- multi-turn task flow

Typical data:
- request-response pairs
- hand-authored exemplars
- rewritten historical examples
- synthetic teacher-generated examples

Typical eval:
- correctness
- policy adherence
- required conventions
- cross-slice consistency

### 3. RL-based preference tuning

Use RL-based preference tuning when multiple outputs are already valid and you want the model to prefer one valid response over another.

Use it for:
- helpfulness ranking
- naturalness ranking
- style or tone preference
- choosing between multiple acceptable completions

Typical data:
- preference pairs or triples
- sampled candidates from the current model
- judge or reviewer labels

Typical eval:
- pairwise win rate
- default-vs-sampled comparison
- preference by slice

## Composition Rule

If the task needs more than one stage, sequence them from broadest to narrowest.

- CPT first when the base distribution is wrong.
- SFT next when the task shape is wrong.
- Preference tuning last when there are several acceptable answers and you want to choose the best one.

## Design Rule

Keep the story simple.
State what behavior you are trying to change, what data supports that change, and how you will know it worked.

From cavaunpeu/agent-skills-closet, MIT.

More from cavaunpeu/agent-skills-closet