Skip to content

Releases

v0.1.0

Choose a tag to compare

@Bobliuuu Bobliuuu released this 13 Nov 08:06

The 10% challenge: our attempt, v0.1.0

A record of our entry to the 10% challenge: shrink a trained model below 10% of its fp16 size, keep its behaviour on an unseen exam, and teach it a new label from at most 25 corrections without training the weights. Task: 7-way voice-skill routing on a locked exam of 2,744 items.

Filed entry: L1 (Bonsai-8B plus a closed-form feedback patch), issue #2. Per the repo's own result files: 7.16% of Qwen3-8B fp16 including the patch, code and calibration prompts; 25 annotations teach the new label to 90.9% (end to end, shipped commands); neighbouring items kept 99.7%; shuffled, wrong and empty feedback teach 0%. It passes the 9 judging rules under our reading; three of those passes rest on readings the maintainer has not ruled on (see README).

C2 artifact (Qwen2.5-7B-Instruct, 9.404% of bf16)

Choose a tag to compare

@Bobliuuu Bobliuuu released this 06 Nov 13:15

Artifact for the proof in DeusExperiments/ten-percent-challenge#2 (DeusExperiments/ten-percent-challenge).

Qwen2.5-7B-Instruct compressed to 9.404% of its bf16 size (1,432,412,701 bytes, 9 files), task C2 (tool selection).

How to verify

  1. Download ALL assets except artifact_MANIFEST.json and memory_combined_k6.json into one empty directory (the artifact directory). Do not add any other file; do not import predict.py from inside it (a __pycache__ would change the size and tree hash).
  2. Check each file against the sha256 values in artifact_MANIFEST.json (field files). The harness tree hash is 6290414d96f3119799516442cab5106243dede6a5b29256f24669f0581ee9661 (computed by v2/harness/common.py tree_sha256; os.walk scheme hash 52dd2397d025f34a85b6905fc844dadd3a1b42b865397041083c7512ec4c642f).
  3. memory_combined_k6.json (852 bytes, sha256 3cc96ec2af0b43ad054868eb6a1b237c8895623bf9a37e880287820ce1793772) is the stage-2 feedback memory; it is NOT part of the artifact directory.
  4. Run: see v2/harness/RUNBOOK.md (needs a GPU, torch, transformers==4.51.3, constriction; decode takes about 3 minutes).

Locked-exam receipts: v2/harness/receipts/REAL_final/stage1.json and stage2.json.