In developmentMoroLM-56M is a research base model, not a chatbot. Things here change often.See what changed
MoroLM
Home / Docs / Project map

Project map

A map of the MoroLM repository, research pipeline and milestone status.

Work in progress: this page describes MoroLM-56M v0.6 and will change as the project moves.

Research pipeline map#

1. Source auditlicense + provenance
2. Cleaningdedupe + quarantine
3. Tokenizer24k frozen UGM
4. Packing512-token blocks
5. Pretrainingnext-token LM
6. Evaluationfixed + external
7. ExportHF → GGUF

Repository map#

atlas-local-ai/
├── tokenizer/
│   ├── atlas_tokenizer.model
│   └── atlas_tokenizer-v0.1.model
├── model/
│   ├── configuration.py
│   └── atlas_model.py
├── training/
│   ├── pretrain_v2.py
│   ├── datasets.py
│   └── checkpoint.py
├── configs/
│   └── morolm56m-base-v0.6-cleanstart.yaml
├── data/
│   ├── staging/
│   │   └── v0.5/train_candidate.jsonl
│   ├── evaluation/fixed/
│   └── evaluation/external/
├── checkpoints/
│   ├── historical v0.1–v0.5 artifacts
│   └── morolm-56m-base-v0.6-cleanstart/
├── models/
│   └── exported HF / GGUF artifacts
├── Go backend (planned/in progress)
└── browser UI (planned/in progress)

Internal paths still preserve the historical Atlas naming. Public documentation uses MoroLM.

Milestone matrix#

AreaStatusNotes
Custom tokenizerWorking24k SentencePiece Unigram, frozen.
56M architectureWorkingLlama-like decoder, GQA, RoPE, RMSNorm, SwiGLU.
Correct causal objectiveVerifiedAlignment smoke + 500-step memorization proof.
Clean base modelv0.612,500 updates from random weights.
Held-out evaluationPartialFixed historical + external sets exist; natural Darija suite is next.
Instruction modelNot yetEarlier SFT probes predate the clean base lineage.
GGUF runtimePipeline provenParity work exists; v0.6 export is a future release step.
Production appFutureGo + llama.cpp + SQLite + streaming web UI target.

Clean lineage rule#

Base checkpoints v0.1 through v0.5 must not be used as parents for the clean MoroLM base lineage because they were trained with the faulty two-token-ahead objective. They remain useful only as historical/debugging artifacts.

↑ ↓ to move · Enter to open · Esc to close