DietRecommendation-Qwen2.5-0.5B

A LoRA adapter for Qwen/Qwen2.5-0.5B-Instruct that turns a short user profile (age, gender, height, weight, activity level, dietary preference, daily calorie target) into a one-day meal plan with breakfast, lunch, snack and dinner.

It is a small, fast, single-purpose model: about 2.2M trainable parameters on top of a 0.5B base, usable on CPU.

Not medical advice. This is an experimental model trained on a small dataset. Its output has not been reviewed by a dietitian and must not be used to manage a health condition. See Limitations.

Model details

Developed by	syubraj
Model type	LoRA adapter (PEFT) for a causal language model
Base model	Qwen/Qwen2.5-0.5B-Instruct
Language	English
License	Apache-2.0
Training data	syubraj/DietRecommendation-dataset-Qwen-2.5-0.5b
Related model	syubraj/DietRecommender_4bit_Qwen2.5-0.5B (trained on the same dataset)

How to use

The adapter was trained on one fixed prompt layout. Use the same system prompt and the same field names and order, otherwise quality drops. Two details matter:

• Pass the system prompt explicitly. Without it, the chat template inserts Qwen's default system prompt, which the adapter never saw.
• The first field is spelled Ages: in the training data, so keep it that way.

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

BASE = "Qwen/Qwen2.5-0.5B-Instruct"
ADAPTER = "syubraj/DietRecommendation-Qwen2.5-0.5B"

tokenizer = AutoTokenizer.from_pretrained(ADAPTER)
model = AutoModelForCausalLM.from_pretrained(BASE, torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(model, ADAPTER)
model.eval()

# Exact system prompt used in training (including the line break).
SYSTEM_PROMPT = (
    "Act as a nutrition expert. Based on the user’s age, gender, height, weight, "
    "activity level, diet preference, and calorie target, \n"
    "suggest a balanced meal plan with breakfast, lunch, snacks, and dinner."
)


def build_profile(age, gender, height_cm, weight_kg, activity, diet, kcal):
    return (
        f"Ages: {age}\n"
        f"Gender: {gender}\n"
        f"Height: {height_cm} cm\n"
        f"Weight: {weight_kg} kg\n"
        f"Activity Level: {activity}\n"
        f"Dietary Preference: {diet}\n"
        f"Daily Calorie Target: {kcal} kcal"
    )


messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": build_profile(32, "Female", 165, 65, "Lightly Active", "Vegetarian", 1600)},
]

inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt", return_dict=True
).to(model.device)

with torch.no_grad():
    output = model.generate(**inputs, max_new_tokens=96, do_sample=False)

print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

device_map="auto" needs accelerate; drop that argument to load on CPU.

Input values seen in training

Field	Values
Gender	Male, Female
Activity Level	Sedentary, Lightly Active, Moderately Active, Very Active
Dietary Preference	Omnivore, Vegetarian, Vegan
Height / Weight	centimetres / kilograms
Daily Calorie Target	kcal

Output format

Four lines, one per meal. This example is a row from the training set and shows the format the model was taught to produce:

Breakfast: Tofu scramble with veggies
Lunch: Lentil soup with whole wheat bread
Snack: Apple with almond butter
Dinner: Vegetable stir-fry with brown rice

Merging the adapter

To ship a standalone model with no PEFT dependency:

merged = model.merge_and_unload()
merged.save_pretrained("DietRecommendation-Qwen2.5-0.5B-merged")
tokenizer.save_pretrained("DietRecommendation-Qwen2.5-0.5B-merged")

Intended uses

• Prototypes and demos of profile-to-meal-plan generation.
• A starting point for meal-idea features where a person reviews the result.
• A worked example of LoRA fine-tuning a very small instruct model on a structured task.

Out of scope

• Medical nutrition therapy, or diet planning for any health condition (diabetes, kidney disease, food allergies, pregnancy, eating disorders and so on). The prompt has no field for conditions, allergies or medication.
• Children and teenagers. The training examples sampled for this card were all adults.
• Any setting where the output is delivered to people as professional advice without qualified review.
• General chat. The adapter is tuned for one prompt format.

Training data

syubraj/DietRecommendation-dataset-Qwen-2.5-0.5b: 1,698 examples (1,358 train / 340 validation), derived from the Kaggle dataset Nutrition Daily Meals in Diseases Cases.

Each example is a single text string already rendered in ChatML (system prompt, user profile, assistant meal plan), 525 to 678 characters long.

Training procedure

LoRA configuration

Setting	Value
Rank (r)	4
lora_alpha	16
lora_dropout	0.1
Bias	none
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Task type	CAUSAL_LM
Trainable parameters	about 2.2M (roughly 0.44% of the base model)

Hyperparameters

Setting	Value
Learning rate	2e-4
Train / eval batch size	4 / 4
Gradient accumulation steps	4 (effective batch size 16)
Optimizer	AdamW (adamw_torch), betas (0.9, 0.999), epsilon 1e-8
LR scheduler	cosine, 100 warmup steps
Epochs	15
Mixed precision	native AMP
Seed	42

Training results

Training loss	Epoch	Step	Validation loss
3.0693	1.18	100	0.2517
0.9641	2.35	200	0.2243
0.8782	3.53	300	0.2218
0.8378	4.71	400	0.2253
0.8114	5.88	500	0.2179
0.7791	7.06	600	0.2193
0.7539	8.24	700	0.2178
0.7247	9.41	800	0.2185
0.6962	10.59	900	0.2234
0.6731	11.76	1000	0.2265
0.6363	12.94	1100	0.2317
0.6184	14.12	1200	0.2343

Final reported validation loss: 0.2343 (step 1200).

How to read these numbers:

• Validation loss bottoms out at step 700 (0.2178) and then climbs slowly while training loss keeps falling. That is mild overfitting; around 8 epochs would have been enough.
• Training and validation loss are on different scales in this log (training is several times higher throughout), so compare each column only with itself.
• The dataset stores each example as one full ChatML string, and every example shares the same system prompt. A loss computed over the whole string is pulled down by that repeated text, so the absolute value says little about meal-plan quality.

Framework versions

• PEFT 0.14.0
• Transformers 4.47.0
• PyTorch 2.5.1+cu121
• Datasets 3.3.1
• Tokenizers 0.21.0

Evaluation

Only validation loss was measured. There is no task-level evaluation yet: no check of nutritional adequacy, calorie accuracy, adherence to the dietary preference, or expert review. Treat the model as unvalidated for all of these.

Limitations

• No quantities. Outputs name dishes only, with no portion sizes, calories or macronutrients. Nothing ensures the plan meets the requested calorie target.
• Weak personalisation. In the training data, the same meal plan appears for different profiles and calorie targets, and identical profiles appear with different plans. Expect the dietary preference to shape the output much more than age, height, weight or calorie target.
• Narrow menu. The training plans draw on a limited, largely Western set of dishes (oatmeal, tofu scramble, grilled chicken salad, lentil soup, salmon with vegetables). Outputs will be repetitive and may not suit other cuisines or budgets.
• Three diet types only. Omnivore, vegetarian and vegan. No support for allergies, intolerances, religious diets, keto, gluten-free and so on.
• Small base model, small dataset. A 0.5B model trained on about 1.4k examples can still produce a non-vegan dish in a vegan plan or other inconsistencies. Validate outputs in code if that matters for your use.
• Possibly optimistic validation loss. The training split contains repeated and near-identical examples; if the validation split overlaps with them, the validation loss understates the real error.
• Fixed format, English only. Behaviour on free-form requests, other languages or other field layouts is untested.

Recommendations

Show a clear "not medical advice" notice wherever outputs reach end users, keep a person in the loop, and check dietary-preference compliance separately rather than trusting the model.

Citation

@misc{syubraj2025dietrecommendation,
  title  = {DietRecommendation-Qwen2.5-0.5B: a LoRA adapter for meal-plan recommendation},
  author = {syubraj},
  year   = {2025},
  url    = {https://huggingface.co/syubraj/DietRecommendation-Qwen2.5-0.5B}
}

Base model:

@misc{qwen2.5,
  title  = {Qwen2.5: A Party of Foundation Models},
  url    = {https://qwenlm.github.io/blog/qwen2.5/},
  author = {Qwen Team},
  month  = {September},
  year   = {2024}
}