Connect Hugging Face NOEON repo to Colab via Git LFS
openhands+1: openhands
fc1c423
- scripts/colab_clone_noeon.py: clone Frankenstein-Labs/NOEON into /content/NOEON
with pointers-only default (GIT_LFS_SKIP_SMUDGE=1), selective --lfs-include,
or --with-weights for the full 43.5 GB. Token (if any) is injected through the
environment only, never written to disk.
- notebooks/NOEON_TRAIN_GPU.ipynb (and its generator): clone cell now uses the
PUBLIC HF repo and skips the LFS smudge; removed the wrong GITHUB_TOKEN/private
assumption.
- docs/COLAB_CLI.md: verified repo facts (public, 380 files, 43.5 GB, LFS),
clone/sync instructions, no-token-leak notes.
Verified: pointer clone = 49 MB / 380 files; selective pull of one checkpoint
matched its LFS SHA-256; 116 tests pass.
Co-authored-by: openhands <openhands@all-hands.dev>
Verify VS Code <-> Colab Remote-SSH bridge
openhands+1: openhands
5295281
- setup_colab_cli.sh now writes the colab-noeon stanza idempotently and installs
the Open VSX Remote-SSH client (jeanp413.open-remote-ssh); MS Remote-SSH is
not on Open VSX.
- docs/COLAB_CLI.md records verified facts: ssh -G parses the host; proxy-mode
without -s never creates a runtime; proxy-mode with -s creates it (verified
live: real SSH banner, session stopped immediately, 0 assignments left).
Co-authored-by: openhands <openhands@all-hands.dev>
Add Google Colab CLI setup: GPU check, setup script, user guide
openhands+1: openhands
628e0f3
- scripts/colab_gpu_check.py: honest GPU/CUDA/torch/VRAM probe, exits 1 with
no CUDA device (verified locally: no GPU, exit 1).
- scripts/setup_colab_cli.sh: idempotent install of the CLI, keyring backend,
OpenSSH client + ed25519 key, and the VS Code Remote-SSH stanza.
- docs/COLAB_CLI.md: measured environment facts, auth flow, Remote-SSH bridge.
No Colab session, GPU, or compute units were created; Google sign-in remains
the user's step.
Co-authored-by: openhands <openhands@all-hands.dev>
Add the central-brain layer and the experiments that justify it, with real,
measured results (no claims that were not run).
- noeon/unified_brain.py: UnifiedBrain facade (representation + memory +
context + comm + planning + control) under one identity.
- noeon/memory/: six-tier memory (working/conversation/episodic/semantic/
long-term/project) with deterministic embeddings, dedup, tier summaries,
controlled update/delete and persistence.
- noeon/context/plan.py: window choice justified by the REAL Qwen2 tokenizer
(max 131072) plus a measured hard-ceiling proof; the fusion core's 8192-token
cap is reported separately, never conflated with the budget.
- noeon/alignment/: real 2560-d extraction via Hub range reads + a contrastive
projection. Measured: the three slots are already aligned at the embedding
level (cos 0.97-0.99); a scrambled control recovers 0.012 -> 0.900, proving
the trainer and metric are real.
- noeon/planning/: deterministic TaskPlanner + ActionValidato…
Add NOEON_LAB JupyterLab workspace with 9 executed research notebooks
openhands+1: openhands
8653213
Set up a real JupyterLab server (4.6.6) on the runtime's exposed port and a
research workspace under notebooks/NOEON_LAB/. The nine notebooks import NOEON's
own modules and run real code: measured parameter counts, real safetensors
headers and tensor rows read straight from the Hub, and real fusion forward
passes. All nine execute with zero cell errors; NOEON's test suite is unchanged
at 90 passed, 1 skipped.
Nothing downloads the 37.8 GB of weights: noeon/lab/hub_range.py reads headers
and leading tensor rows via HTTP range requests, so a full run uses ~66 MB of
disk. Two real bugs surfaced while running the notebooks in the browser and were
fixed: shutil.sysconf -> os.sysconf, and safetensors data_offsets are relative
to the data section (8 + header_len), not the file start.
Adds scripts/start_jupyterlab.sh (token read from the environment, never written
to disk), scripts/run_noeon_lab.py, scripts/make_noeon_lab_notebooks.py,
requirements-lab.txt and docs/NOEON_LAB_JUPYTERLAB.md.
…
Import all source weights into NOEON + loading tests
openhands+1: openhands
c0dd269
Import the full frozen source weights of the three models into
weights/upstream/<slot>/ (119 files, 37.803 GB): microsoft/Mage-VL (vision),
microsoft/Mage-VL-Cowork (computer_use), mage-flow-community/Mage-Flow
(generation). Every file was downloaded, SHA-256 checked against the recorded
manifest, uploaded, then freed; the import is resumable and disk-bounded.
Add scripts/import_sources_to_hf.py and scripts/test_loading_from_hub.py, the
import report (weights/import_report.json), the loading-test report
(reports/loading_tests.json) and docs/IMPORT_REPORT.md.
Verification (from the Hub, not local state): 119/119 files present, size and
LFS oid (== SHA-256) match; two shards re-downloaded and re-hashed. Loading:
all shard headers read (vision 760 tensors, computer_use 696, generation VAE
839 + transformer 397), tokenizer and config load for vision/computer_use;
full from_pretrained skipped (needs ~23 GB RAM, box has 15 GB). Update WEIGHTS,
SOURCES, MANIFEST, STATUS, DELIVERY_REPORT, RE…
NOEON: real context system, inter-model communication, integrated brain
openhands+1: openhands
3f71e8c
Adds the three layers that turn NOEON from an architecture into a system,
plus the evidence from the real source weights.
- noeon/context/ conversation/working/task/long-term memory, BM25(+dense)
retrieval, hard context budget with extractive compression, session
persistence. 14 tests. docs/CONTEXT.md. States clearly that external
memory never enlarges a model's native window.
- noeon/comm/ standardised Message envelope, bus with deadlines/retries/
exception containment, result verification, multi-step coordinator with
shared context, peer help requests and capability reservation. External
tools + a deterministic planner on the same bus. 22 tests.
docs/COMMUNICATION.md.
- noeon/brain.py integrated system (architecture + context + comm). 5
end-to-end tests.
- scripts/ download_and_verify_sources.py (streaming SHA-256 + tensor
inventory, disk-bounded), verify_real_bytes.py (bridges on real Hub bytes
via range requests), run_noeon_demo.py (end-to-end demo).
- weig…
NOEON: official identity (Frankenstein-Labs) + premium model card; proprietary-weights structure