01
Closed corpus
Training text comes from glossary, grammar, story facts, and admitted memories named in the ledger.
By Eidetic Labs · Closed-world language model
Big idea. Tiny package.
QuarkLM is a research prototype for language models that learn only from their admitted dataset. No pretrained weights. No pretrained tokenizer. No external embeddings. Just a tiny world, exact retrieval memory, a corpus-trained tokenizer, guarded weight updates, generated probes, and evidence that follows every promotion.
Experimental scaffold. Not a production assistant. The point is auditable growth from a closed corpus.
A small world with hard edges
Instead of training on everything, train only on what was admitted.
Large language models usually compress broad mixed corpora into weights, then specialize that base with retrieval, prompting, fine-tuning, or adapters. QuarkLM reverses the order: a new lesson enters the corpus, retrieval memory can serve it immediately, training candidates decide whether weights should learn from it, and evaluation accepts or rejects the guarded update.
The model starts small on purpose.
The prototype is a tiny stack of corpus generation, reliable response logic, learned answer selection, generative answer decoding, and an experimental from-scratch transformer.
02
Random weights
Learned components start from random initialization and write versioned checkpoints under runs.
03
Operational self
The model can know its dataset boundary, update process, and unknown policy without claiming subjectivity.
04
Own tokenizer
The current tokenizer learns characters from admitted text; future tokenizers must be corpus-derived too.
Self-improvement means measured change.
QuarkLM does not silently mutate. Its loop is explicit: new lesson -> corpus -> retrieval memory -> training candidates -> guarded weight update -> evaluation -> accepted or rejected.
- 01
Corpus
New knowledge enters through admitted, ledgered corpus files.
- 02
Memory
Exact retrieval serves admitted knowledge before weights claim it.
- 03
Candidates
Training examples are built from admitted sources and failure reports.
- 04
Update
Transformer weights receive guarded consolidation pressure.
- 05
Evaluate
Coverage, provenance, forgetting, and branch-diversity gates decide acceptance.
- 06
Publish
Docs and marketing update when they reference current state.
Promoted evidence: v0.42. Latest screen: v0.115.0.
The latest promoted self-improvement run is `runs/self-improve-v0.42/`. The latest transformer screen, v0.115.0, tests a bias-frozen hidden-projection margin candidate. It reduces collapsed-token hidden advantage, but still keeps branch diversity as the promotion blocker.
RC postureprototype near
Research Prototype RC is the near-term track; Language Model RC waits on branch routing.
Admission probes48/48 · 84/84
Direct and paraphrase probes generated from admitted memory.
Glossary probes38/38
Definition probes generated from admitted glossary words.
Projection repair0.0842 → 0.0736
v0.115.0 lowers collapsed-token hidden advantage, but 9/9 branch profiles still collapse.
Self-diagnosisno external model
Rule-based report diagnosis found zero blockers on v0.42.
No-candidate generator219/219
v0.31 auxiliary generator still proves exact answers without a candidate set.
Promotion gatepassed
Exact eval, forgetting, probe, and leakage audits all passed.
Where QuarkLM goes next.
Package migration
Move the Python import path from `closed_world_lm` toward QuarkLM naming without breaking run artifacts.
Release-synced docs
Keep technical docs and this marketing page aligned with promoted evidence.
Larger learning batches
Admit more memory while preserving anti-forgetting checks and generated eval coverage.
Release candidate boundary
Package an honest Research Prototype RC first, then resume transformer work as a profile-balanced routing repair without relaxing promotion gates.
QuarkLM · research prototype