- Python 80%
- HTML 19.6%
- Shell 0.2%
- CSS 0.1%
- Astro 0.1%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
|
All checks were successful
checks / select relevant integrity checks (push) Successful in 11s
checks / E01 candidate surface sufficiency (push) Has been skipped
checks / DMS-003 protocol freeze integrity (push) Has been skipped
checks / DMS-003 CPU successor integrity (push) Has been skipped
checks / resource accounting and workflow consistency (push) Has been skipped
checks / documentation integrity (push) Successful in 10s
|
||
| .github | ||
| agents | ||
| configs | ||
| continuum_research | ||
| docs | ||
| experiments | ||
| reports | ||
| research | ||
| results | ||
| scripts | ||
| site | ||
| src/smolmodelcompany | ||
| templates | ||
| tests | ||
| .gitattributes | ||
| .gitignore | ||
| AGENTS.md | ||
| LICENSE | ||
| pyproject.toml | ||
| README.md | ||
| WEEKEND_RESEARCH_BRIEF.md | ||
The Smol Model Company
Research small models that reason, compose skills and adapt, with reproducible evidence and practical resource limits. Transfer mechanisms that work into a useful local assistant.
Public site: thesmolmodel.co. Canonical work: Forgejo.
Working here
Read AGENTS.md and the short research contract. Cassie approves a research question and bounded resources once; agents execute, keep evidence and report. Agents own JSON/configs/receipts. Ordinary code/docs use relevant tests, a PR and plain-language approval rather than research-admission paperwork. See workflow and compute.
Research and evidence
The experience-consolidation program records a research direction; task-specific current instructions determine the next experiment. The evaluation principles guide fair comparisons, held-out evaluation and honest claims. Results are generated from retained run reports.
Continuum preserves the frozen synthetic adaptation study, including its negative results, source/config identities, reproduction checks and separate private-evidence handoff. See its migration record for provenance and follow-up.
Earlier research designs, backlog assignments, decisions and experiment protocols remain evidence of their time, not additional active governance. Preserve them and their corrections. GitHub is a suspended mirror, not a second work queue.
Keep exact code/config/data/model identities and reproduction commands with measurements, failures, costs and limitations. Keep large weights/datasets and private information outside Git. The MIT license does not relicense third-party artifacts. No new empirical execution or spending is authorized by this reset.