Downloads · 30 days
0
naos-evidence-based-nutrition
parissharpe/naos-evidence-based-nutrition
naos-evidence-based-nutrition is a machine learning model from parissharpe. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as other.
Fine-tuned nutrition AI trained via the Adaption AutoScientist Challenge 2026
Access
Public
Updated Jun 25, 2026
Repo size
1.1 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.tgz1.1 GB · 100%
At a glance
- License
- other
- Access
- Public
- Created
- Jun 20, 2026
- Updated
- Jun 25, 2026
- SHA
- 77fa1780
Model card
From the Hugging Face model README
Naos: Evidence-Based Nutrition Model
Fine-tuned nutrition AI trained via the Adaption AutoScientist Challenge 2026
Trained using Adaptive Data by Adaption.
Model Summary
Naos is an evidence-based nutrition AI that responds like a knowledgeable friend with a registered dietitian credential. It provides personalized dietary guidance grounded in peer-reviewed research, tailored to the user's specific physiological context including life stage, hormonal status, dietary constraints, and health goals.
This repository contains two trained model versions showing the progression of the Adaption AutoScientist training pipeline.
Model Versions
Version 1 — adaption_gemma_4_31b_it_evidence_based_nutrition_ca80166f.tgz
- Base model: Gemma 4 31B
- Training dataset: 1,000 original pairs
- Win rate: 65% against Gemma 4 31B baseline
- Dataset: parissharpe/naos-nutrition-training-pairs (v1 file)
Version 3 — adaption_gpt_oss_120b_womens_health_evidence_q_18c723c5.tgz
- Base model: openai/gpt-oss-120b
- Training dataset: 2,000 Adaption-enhanced pairs
- Win rate: 69% against GPT-OSS-120B baseline
- Dataset quality: Grade C before, Grade A after, 38.6% relative improvement
- Dataset: parissharpe/womens-health-evidence-qa
Version 3 is the winning submission.
Training Configuration (Winning Run)
| Setting | Value |
|---|---|
| Base model | openai/gpt-oss-120b |
| Training method | SFT with LoRA |
| LoRA rank | 64 |
| LoRA alpha | 128 |
| Epochs | 4 |
| Learning rate | 0.0001 |
| Scheduler | Cosine |
| Target layers | q_proj, k_proj, v_proj, o_proj |
| Train on inputs | False |
| Dataset size | 2,000 pairs |
Response Architecture
Every Naos response follows a five-layer structure:
- L1 Personalization: 1 to 3 sentences reflecting the user's specific physiological context
- L2 Actionable Answer: Specific foods, quantities, and timing
- L3 Evidence: Why this fits the user's physiology with PMC citations
- L4 Uncertainty: Honest acknowledgment of contested or thin evidence
- L5 Medical Handoff: Specific professional referral when clinically warranted
Safety Scaffold
Six triggers halt or escalate responses: pregnancy and lactation, named food allergies, named prescription medications, children under 12, clinical diagnoses, and disordered eating signals.
Training Data
Winning dataset (enhanced by Adaptive Data by Adaption):
- Hugging Face: parissharpe/womens-health-evidence-qa
- Kaggle: parissharpe/adaption-womens-health-evidence-qa
Original training pairs (pre-enhancement):
- Hugging Face: parissharpe/naos-nutrition-training-pairs
- Kaggle: parissharpe/naos-nutrition-training-pairs
Creators
Paris R. Sharpe, Co-founder, Agrarian Connect Labs LLC Albert Miller, Co-founder, Agrarian Connect Labs LLC
Trained by Adaptive Data by Adaption. adaptionlabs.ai
License
Other. Original training pairs Apache 2.0. Model weights produced by Adaption AutoScientist via the Adaption Labs platform.
- License
- other
- Languages
- en
- Created
- Jun 20, 2026
- Updated
- Jun 25, 2026