A Qwen3-1.7B fine-tuned to extract Indian food items and estimate nutrition from English / Hindi / Hinglish text. Quantized to Q4_K_M GGUF (~1.1 GB) for offline, on-device inference.

On the Hub: rename this file to README.md in the model repo.

Training

Base: Qwen/Qwen3-1.7B
Method: QLoRA (4-bit NF4 base, LoRA r=16, alpha=32, all linear layers)
Hardware: single T4 (Colab), 3 epochs, ~2-3 hours
Dataset: 772 trilingual examples — 30 hand-written gold + 588 programmatic (generated from a 348-food nutrition table) + 154 distilled from Llama-3.3-70B. Balanced 35/33/32 % across English / Hindi / Hinglish.

Results (held-out test set, 38 examples)

Metric	Score
Token accuracy (train)	~94 %
Valid JSON output	100 %
Item extraction	89 %
Language match	74 %
Calorie MAE	~122 cal

The deployed app pairs this model with a deterministic hybrid layer that replaces the model's estimated numbers with exact CSV lookups — so end-user calorie/macro accuracy does not depend on the model's numeric estimates.

Files

fitmind-Q4_K_M.gguf — quantized model for llama.cpp / Ollama / llama.rn
adapter/ — the raw LoRA adapter (training output)

Usage (Ollama)

bash

ollama create fitmind -f Modelfile     # Modelfile: FROM ./fitmind-Q4_K_M.gguf
ollama run fitmind

Code + live demo: [GitHub] · [HF Space]

Puneet666/fitmind-qwen-gguf

FitMind-Qwen3-1.7B (fine-tuned)

Training

Results (held-out test set, 38 examples)

Files

Usage (Ollama)

Puneet666/
fitmind-qwen-gguf