brixton/openlens-style-lab
2
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity
OpenLens style/OPSD research lab: datasets, evals, reports, and BRE-412 parser handoff.
53 commits 1 branch 0 tags 39 MiB
  • HTML 52.3%
  • Python 47.5%
  • Shell 0.2%
Find a file
Exact
Brixton Leigh a3bf1df258 Add animated OPSD narrated deck video
2026-06-12 18:06:02 -07:00
artifacts/videos Add animated OPSD narrated deck video 2026-06-12 18:06:02 -07:00
coordination Add source-code reward anchor and coordinator state 2026-05-28 02:21:25 -07:00
data Add OpenCode MCP next-tool KL repair v0 2026-06-04 07:33:12 -07:00
docs Store small paper PDFs directly for Forgejo 2026-06-11 11:12:13 -07:00
experiments Record final KL repair training metrics 2026-06-04 07:35:36 -07:00
references/papers Store small paper PDFs directly for Forgejo 2026-06-11 11:12:13 -07:00
scripts Add OpenCode MCP next-tool KL repair v0 2026-06-04 07:33:12 -07:00
tests/fixtures Add OpenCode OPSD rewrite target scaffold 2026-06-03 06:29:48 -07:00
.gitattributes Add OPSD narrated deck video 2026-06-12 17:43:19 -07:00
.gitignore Add OpenLens data and eval scaffolding 2026-05-27 05:20:43 -07:00
program.md Add OpenLens iteration 2 handoff 2026-05-27 21:17:57 -07:00
README.md Add OpenLens paper research synthesis 2026-05-27 05:25:59 -07:00

README.md

openlens-style-lab

Experiment workspace for Bread Technologies / OpenLens writing-style self-distillation on Tinker.

The target is not a generic Brixton identity model. The target is an AI that can draft extremely high-quality externally-facing content for Bread Technologies, Inc., now rebranding / DBA as OpenLens.

Objective

Train and evaluate models that can write in the OpenLens/Bread editorial style:

  • rigorous, evidence-aware AI visibility analysis
  • agency-native positioning
  • precise distinction between public evidence, internal measurement, and opinion
  • strong externally-facing prose suitable for blog posts, landing pages, reports, launch copy, and customer-facing explainers

Inputs

Seed source manifest:

data/source_manifest.jsonl

Reference repos:

/Users/brixton/repos/bread-pg
/Users/brixton/repos/autoresearch
/Users/brixton/repos/tinker-lab
/Users/brixton/repos/tinker-cookbook

Constraints

  • Total user-approved spend cap: USD 1000 across Tinker and OpenRouter for this project.
  • Do not commit API keys, raw browser cookies, account tokens, or logs that may contain secrets.
  • Do not rely on a single generated training run without a held-out eval set.
  • Do not train a model to fabricate market statistics. Preserve the OpenLens habit of dating claims and distinguishing measured results from public-record synthesis.

Initial Model Ladder

  1. Smoke: meta-llama/Llama-3.2-1B, meta-llama/Llama-3.2-3B, or Qwen/Qwen3-4B-Instruct-2507
  2. Sweet spot: Qwen/Qwen3-30B-A3B, Qwen/Qwen3.5-27B, Qwen/Qwen3.5-35B-A3B, Qwen/Qwen3.6-35B-A3B, or nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
  3. Expensive/final: best available Kimi teacher/student path, currently candidates include moonshotai/Kimi-K2.6, moonshotai/Kimi-K2.6:peft:131072, moonshotai/Kimi-K2.5, and moonshotai/Kimi-K2-Thinking

Check live capabilities first:

cd /Users/brixton/repos/tinker-lab
uv run python scripts/list_capabilities.py

Recommended Next Command

Read the project brief, coordinator prompt, and paper synthesis:

sed -n '1,260p' docs/project_brief.md
sed -n '1,260p' docs/coordinator_prompt.md
sed -n '1,260p' references/papers/notes/synthesis.md

Then run the zero-spend data/eval scaffold before spending Tinker or OpenRouter budget:

python3 scripts/collect_sources.py --force
python3 scripts/validate_eval_set.py --evals data/evals/openlens_style_holdout_v1.jsonl