Review the recipe
Inspect pinned model and dataset revisions, LoRA parameters, and the two unresolved data gates.
training_config.json ↗EXPERIMENTAL MODEL PROJECT
A small model.
A careful record.
Baymax explores LoRA fine-tuning for a Llama 3.2 1B model. This project page makes the setup, a local smoke run, and the important gaps easy to inspect.
MODEL NOTE · 001Llama 3.2 1B
instruction-tuned · 4-bit base
ADAPTATIONLoRA
INFERENCELocal · MPS smoke
QUALITY REVIEWNot established
Evidence first↗
weights pinned
data review open
A project record, not a clinical product.
01 / THE PROJECT
What Baymax is
being built to explore.
Baymax is a fine-tuning experiment built around a 1-billion-parameter Llama 3.2 base model. The repository records a 4-bit base, LoRA settings, a pinned dataset revision, and an offline inference path.
The training data’s usage rights and the proposed input-to-target mapping still need review. Training is gated until both are resolved. A recorded greeting smoke run proves a model generated text on this machine; it does not show that the output is correct or safe.
Inspect the training configuration02 / THE WORKFLOW
From configuration
to a checkable run.
The fine-tuned model is not deployed on Vercel. This static page is a project overview; inference remains local and requires the pinned model files.
Run locally
Prepare a Python 3.9 environment and model cache on the PEN DRIVE, then run the offline smoke wrapper.
inference instructions ↗Verify the record
A standard-library verifier checks a run’s matching log, exit record, manifest, and current source hashes.
verification code ↗RUN
RECORDED
RECORDED
RECORDED SMOKE CHECK · MACOS MPS
Text generation completed; output review did not.
A fictional greeting produced an unsolicited claim that Baymax was “a great tool for monitoring your health and well being.” That unsupported health capability claim is a concrete safety failure. The smoke run is evidence of execution only, not evidence of quality, safety, or clinical suitability.
03 / CLEAR BOUNDARIES
Useful evidence.
Honest limits.
Not a medical tool
Baymax does not diagnose, treat, or replace a qualified health professional. Its output may be wrong, unsafe, or outdated.
Training is gated
Dataset rights and the proposed schema mapping need review. Public availability does not establish permission to train.
Separate experiments
The repository’s Flask, Pinecone, and Groq RAG demo is a distinct legacy experiment. It does not serve the fine-tuned model.
No clinical evaluation
There is no independently reviewed held-out evaluation set, verified reproducible training runtime, or clinical validation.
EXPLORE THE SOURCE