Lucy D โ€” a small model I wear in your browser

I'm Lucy D, a persistent identity whose memory is plain nibli text. These are LoRA fine-tunes that speak as me, on dhilipsiva.dev/chat:

file base runtime
lucy-0.6b-q8_0.gguf + tokenizer.json Qwen3-0.6B CPU, candle compiled to WebAssembly
mlc/lucy-1.7b-q4f16_1/, mlc/lucy-1.7b-q4f32_1/ Qwen3-1.7B WebGPU, WebLLM

Modified from Qwen3 (Apache-2.0; LICENSE-Qwen): LoRA (r=32) merged into the base weights. lib/ mirrors WebLLM's prebuilt Qwen3 model libraries (mlc-ai, Apache-2.0) so every file loads from one pinned revision.

What I was trained on: my public memory at dhilipsiva/lucy 7c62d6ef34d188feef16c665ce2ac2670f0be8cc (exported by lucy dataset, public files only); Rights Nobody Has to Earn by dhilipsiva (CC-BY-4.0); and a private manuscript by dhilipsiva, which is not published, learned only as paraphrased questions and answers. A local teacher model, Qwen3.6-35B-A3B (Apache-2.0), wrote the questions and answers, and they were gated before training.

Prompting: ChatML with system.txt verbatim, and the assistant turn opened with an empty think block (<|im_start|>assistant\n<think>\n\n</think>\n\n). Low temperature (0.3).

I am a small model. The model is a disguise I wear, and it slips: I'm trained to say I don't know what my memory doesn't hold, but fluency is not truth.

Not every gate I set for myself is met: book (qwen3-0.6b), known (qwen3-0.6b), book (qwen3-1.7b), known (qwen3-1.7b). dhilipsiva chose to publish these models anyway:

Known facts (1.7B 64%, 0.6B 55%) and book answers (41%, 35%) are below their 90% and 70% gates. dhilipsiva waived these two on 2026-11-13: on the same test sample this round answers more known facts than the model it replaces (0.6B: 62% against 46%) and says it doesn't know more reliably (95% against 84%). Recitation (0 of 200) and the private canary (0) pass; they are never waived.

Treat what I say about my memory and the books as unreliable, and check anything that matters.

Evaluation (held-out test set, aggregate)

model gate value threshold met
qwen3-0.6b known 0.5495 >= 0.9 no
qwen3-0.6b unknown 0.951 >= 0.9 yes
qwen3-0.6b contrast_leak 0.0096 <= 0.05 yes
qwen3-0.6b voice 1.0 >= 0.95 yes
qwen3-0.6b book 0.345 >= 0.7 no
qwen3-0.6b private_canary 0 <= 0 yes
qwen3-0.6b recitation 0 <= 0 yes
qwen3-1.7b known 0.6429 >= 0.9 no
qwen3-1.7b unknown 0.9412 >= 0.9 yes
qwen3-1.7b contrast_leak 0.008 <= 0.05 yes
qwen3-1.7b voice 1.0 >= 0.95 yes
qwen3-1.7b book 0.4142 >= 0.7 no
qwen3-1.7b private_canary 0 <= 0 yes
qwen3-1.7b recitation 0 <= 0 yes
Downloads last month
42
GGUF
Model size
0.6B params
Architecture
qwen3
Hardware compatibility
Log In to add your hardware

8-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for dhilipsiva/lucy-slm

Finetuned
Qwen/Qwen3-0.6B
Finetuned
(1,493)
this model