Bootstrap self-improving

The self-teacher loop is the most interesting version of Sprint 43: your current adapter writes new ::instruction:: sections for its own document, then the next train run folds them back in.

This is not magic. It works because DLM already has:

• replay-backed retraining
• synthesized instruction provenance (auto_synth)
• a local sway judge for filtering weak candidates

Used carefully, it turns one trained document into a steadily better instruction corpus.

The honest starting point

--teacher self uses the current adapter for that .dlm. That means the loop starts after there is already a trainable local adapter.

A good bootstrap pattern is:

1. Start with prose plus at least some useful seed supervision, or do an initial train from prose and existing sections.
2. Run dlm synth instructions --teacher self.
3. Retrain on the accepted synth sections.
4. Repeat in small batches.

If the adapter still cannot answer basic questions about the document, synthetic instruction generation will mostly amplify noise.

Minimal loop

Train once:

uv run dlm train notes.dlm

Generate a small accepted batch from the current adapter and write it back immediately:

uv run dlm synth instructions notes.dlm \
  --teacher self \
  --per-section 1 \
  --strategy extraction \
  --max-pairs 4 \
  --apply

Retrain on the expanded instruction set:

uv run dlm train notes.dlm

Then inspect real output quality:

uv run dlm prompt notes.dlm "What does DGEMM do?"

That is the basic self-improving loop.

Safer staged version

If you want to inspect before writing:

uv run dlm synth instructions notes.dlm \
  --teacher self \
  --per-section 1 \
  --strategy extraction

uv run dlm synth list notes.dlm

The current implementation stages accepted synth sections for inspection, but it does not yet have a separate dlm synth apply subcommand. Use --apply on the synth run when you want the sections written straight into the document.

Why sway stays the default

The self-teacher path is the place where the default --filter sway matters most.

Without filtering, a weak adapter can happily generate:

• duplicates
• overly generic answers
• plausible but wrong extrapolations

The current synth filter stack is:

1. dedup
2. optional judge pass
3. optional threshold cut

The CLI prints those counts so you can tell whether the loop is getting better or just louder.

A conservative rhythm

This is a healthy local rhythm for a real project:

uv run dlm train notes.dlm
uv run dlm synth instructions notes.dlm \
  --teacher self \
  --per-section 1 \
  --max-pairs 4 \
  --apply
uv run dlm train notes.dlm
uv run dlm prompt notes.dlm "Explain the core idea."

Keep the accepted batch small at first. The point is to improve the document's instruction surface, not flood it with speculative rows.

When to switch away from self

The self-teacher is convenient, but not always the right teacher.

Prefer an external teacher when:

• the local adapter is still very early and weak
• you need broader general knowledge than the current adapter can supply
• you want to compare local-vs-external synth quality on the same prose

That usually looks like:

uv run dlm synth instructions notes.dlm \
  --teacher hf:Qwen/Qwen2.5-1.5B-Instruct \
  --per-section 1 \
  --apply

and then later moving back to --teacher self once the adapter has real domain traction.

Pairing Sprint 43 with Sprint 42

Instruction synthesis and preference mining are complementary:

• dlm synth instructions grows the SFT side of the document
• dlm synth preferences / dlm preference mine sharpens ranking and behavior once the adapter can already produce multiple plausible answers

A practical sequence is:

1. train
2. synth instructions
3. train
4. mine preferences
5. train preference phase

That is the closest current DLM path to a fully local self-improving document loop.

Failure modes to watch

The second pass is not better

That usually means one of:

• the first synth batch was too weak
• the document still lacks enough domain prose
• the adapter is too small for the domain

Do not assume "more synthetic rows" automatically means "better model."

Expansion mode gets weird

--strategy expansion is useful, but it is also the fastest route to polished nonsense. Prefer extraction for early loops and only widen to both or expansion once the adapter is already grounded.

Prompt quality improves but factuality does not

That is a signal to go back to better prose or hand-authored instructional supervision. Self-improvement cannot invent missing source knowledge.

See also

• Synthesize training data
• Instruction section reference
• Self-improving loop
• Reward-model integration