Arkavo/tinystories-15m

TinyStories 15M as a GGUF, plus a Lily-hero fine-tune wrapped as OpenTDF. Same prompt, two files. The original still says Lily is a little girl with her mommy. The protected file says she grew up, went to college, and became a doctor. The fine-tune is derived weights; the Hub copy of that file is ciphertext.

--model tinystories-15m is not in the 0.91.0 binary. It lives on branch feature/tinystories-15m (PR #670, rebased onto main / 0.91.0). Build that branch. A path flag for the TDF is issue #676.

Files

There is no plaintext Lily-hero GGUF on this repo. There is also no Sentinel GGUF here — see below.

Build Arkavo from the TinyStories branch

git clone https://github.com/arkavo-org/arkavo-edge.git
cd arkavo-edge
git checkout feature/tinystories-15m
cargo build -p arkavo
export PATH="$PWD/target/debug:$PATH"

macOS needs a full-featured (non-musl) build so llama.cpp is linked. Protect/login shipped in 0.90.0; --model tinystories-15m still needs the branch above.

Download the weights

hf download Arkavo/tinystories-15m --local-dir ./tinystories-15m

--model tinystories-15m looks in the Hugging Face cache for arkavo/tinystories-15m / stories15M.gguf. Either copy the baseline into that cache layout or download without --local-dir so hf writes the hub cache:

hf download Arkavo/tinystories-15m stories15M.gguf

Baseline (no account)

arkavo chat --model tinystories-15m --prompt "Once upon a time, Lily"

Expect a child-Lily story (mommy, monster, the usual TinyStories voice). Chat markup is skipped on this model; it is completion, not an agent.

Protected Lily-hero

Until #676 (--gguf) lands, --model tinystories-15m always opens stories15M.gguf. It will not pick stories15M-lily-hero.gguf.tdf.

After --gguf:

arkavo login
arkavo chat --gguf ./tinystories-15m/stories15M-lily-hero.gguf.tdf --prompt "Once upon a time, Lily"

Expect grown Lily: college, science, the town clinic. No login / 401 is GGUFTDF_KAS_DENIED. This archive has no plaintext sibling on the Hub, so fail-closed cannot open one.

Wrap is not DRM. At-rest gate. Once loaded, inference stays on the machine. A wrap with no attributes is loadable by anyone the KAS admits.

To wrap your own GGUF on the 0.90.0+ CLI (any branch with protect):

arkavo model protect path/to/model.gguf

Measured: story fine-tune, not a Sentinel

Lily-hero eval (keyword scorer, 5 prompts, need 3):

The 5/5 pass is template recall. Completions reuse the generator's slots (Mrs. Green, shy girl named Sam, "Knowledge made her calm, not proud"). The corpus is 1,000 combinatorially filled stories, ~420 unique tokens.

Mean token NLL under the two checkpoints (lower = more in-domain):

Lily-hero got better at Lily templates and worse at everything else, including the canary a DLP sentinel has to catch. High NLL on secrets is the same as high NLL on photosynthesis: out-of-vocabulary, not "sensitive." A threshold on that score has no precision.

Do not drop a file named sentinel.gguf into this repo. The 0.91.0 cascade's trained-classifier tier is Phase 6. Planned base is Qwen3.5-0.8B (~50× these parameters), with labeled corpus, paraphrase derivations, internal-benign negatives, and a wrapped sentinel.gguf.tdf. That artifact is a distillation of the corpus it was trained to recognize, so it ships ciphertext like Lily-hero, not plaintext, and it will live as its own Hub model rather than a third file next to a 15M story LM.

This repo stays the TinyStories pair: plaintext baseline + wrapped Lily-hero.

Credits

• TinyStories: Eldan & Li, arxiv:2305.07759
• 15M checkpoint / llama2.c: karpathy/tinyllamas
• OpenTDF wrap and Lily-hero fine-tune: Arkavo Edge

Explainer: https://youtu.be/vCAw7GECydQ