README.md

---
base_model:
- joackimagno/Qwen-2.5-General-Recipe-Generation
- Qwen/Qwen2.5-7B
tags:
- text-generation-inference
- transformers
- unsloth
- qwen2
license: apache-2.0
language:
- en
datasets:
- joackimagno/general-recipes
- joackimagno/FILIPINO_RECIPES_2K_V2
metrics:
- bleu
- rouge
- meteor

model-index:
- name: MASID-v1.2
  results:
  - task:
      name: Text Generation
      type: text-generation
    dataset:
      name: joackimagno/FILIPINO_RECIPES_2K_V2
      type: joackimagno/FILIPINO_RECIPES_2K_V2
      split: test
    metrics:
    - name: BLEU-4
      type: bleu
      value: 0.07
    - name: METEOR
      type: meteor
      value: 0.35
    - name: ROUGE-L (F1)
      type: rouge
      value: 0.32
      unit: f1
      config: rougeL
---

# MASID-v1.2

**MASID-v1.2** is a transfer-learned Filipino main-dish recipe generator.  
It is trained **on top of** the base model **[`joackimagno/Qwen-2.5-General-Recipe-Generation`](https://huggingface.co/joackimagno/Qwen-2.5-General-Recipe-Generation)**, which itself was fine-tuned from **Qwen2.5-7B** using **~60k** general recipes from **`joackimagno/general-recipes`**.  
**MASID-v1.2** then performs **a second-stage fine-tuning** on **`joackimagno/FILIPINO_RECIPES_2K_V2` (~2k)** to specialize in **Filipino main dish generation**.

The goal is to generate structured and culturally faithful Filipino recipes while benefiting from broader cooking knowledge learned during the general-recipe stage.

---

## Model Details
- **Base Model (stage 0)**: [Qwen/Qwen2.5-7B](https://huggingface.co/Qwen/Qwen2.5-7B)  
- **Intermediate Model (stage 1)**: [`joackimagno/Qwen-2.5-General-Recipe-Generation`](https://huggingface.co/joackimagno/Qwen-2.5-General-Recipe-Generation) — trained on ~60k general recipes  
- **Specialization Dataset (stage 2)**: [`joackimagno/FILIPINO_RECIPES_2K_V2`](https://huggingface.co/datasets/joackimagno/FILIPINO_RECIPES_2K_V2) (~2,000 samples)  
- **Objective**: Recipe text generation (Filipino cuisine, main dishes)  
- **Method**: Transfer learning (continued fine-tuning from the general-recipe model)

---

## Intended Use
- Assisting in **recipe writing**
- Exploring **Filipino food culture**
- Generating **cooking instructions** in natural language

---

## Limitations
- Trained on a relatively **small Filipino dataset (~2k)** for the specialization stage.  
- May occasionally produce **hallucinated ingredients** or **imprecise steps**.  
- Not a substitute for **nutrition** or **food-safety** advice.  
- Best for **research, education, and creative** use cases.

---

## Evaluation

| Dataset                            | Split | BLEU-4 | METEOR | ROUGE-L (F1) |
|------------------------------------|:-----:|:------:|:------:|:------------:|
| joackimagno/FILIPINO_RECIPES_2K_V2 | test  |  0.07  |  0.35  |    0.32      |

> Notes: Evaluated with Alpaca-style prompting; simple post-processing (strip, EOS truncation).  
> If you rerun evaluation, pin dataset and package versions for reproducibility.

## Dataset Comparison:
  | Dataset                            | Description |
  |------------------------------------|:------------:|
  | joackimagno/FILIPINO_RECIPES_2K| Ingredient Name excludes basic pantry items (e.g. oil, water) but includes any ingredients|
  | joackimagno/FILIPINO_RECIPES_2K_V2 | Ingredient Name only contains classified ingredients from the small object detection model|

---

---

This Qwen2 model was trained **2× faster** with [Unsloth](https://github.com/unslothai/unsloth) and Hugging Face’s TRL library.

[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)

## Example Usage

```python
from typing import List
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig

# Load model and tokenizer
model_name = "joackimagno/MASID-v1.2"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16,
    device_map="auto",
)

# ==============================================================
# Alpaca-style prompt
# ==============================================================

SYSTEM_INSTRUCTION = (
    "You are a Filipino chef. Generate Filipino MAIN DISH recipes.\n"
    "Follow these output rules:\n"
    "1) Use standard stovetop or oven methods.\n"
    "2) Keep steps concise and logically ordered.\n"
    "3) Output FORMAT and ORDER must be exactly:\n"
    "   Recipe name, Prep time, Cook time, Total time, Servings,\n"
    "   Full Ingredients (numbered list), Instructions (numbered list)"
)

ALPACA_TEMPLATE = (
    "Below is an instruction that describes a task, paired with an input that "
    "provides further context. Write a response that appropriately completes the request.\n\n"
    "### Instruction:\n{}\n\n### Input:\n{}\n\n### Response:\n{}"
)

def make_model_input_from_ing(ing_names: List[str]) -> str:
    return (
        "Ingredients to use: " + ", ".join(ing_names) + ".\n"
        "Task: create a Filipino main dish recipe using these ingredients. "
        "Keep steps concise, clear, and coherent."
    )

# Example input
ing_names = ["Beef", "Potato", "Sili", "Carrot", "Sayote"]

alpaca_prompt = ALPACA_TEMPLATE.format(
    SYSTEM_INSTRUCTION,
    make_model_input_from_ing(ing_names),
    ""  # leave response empty for model to generate
)

# ==============================================================
# Run inference
# ==============================================================

inputs = tokenizer(alpaca_prompt, return_tensors="pt").to(model.device)

gen_config = GenerationConfig(
    max_new_tokens=512,
    temperature=0.7,
    top_p=0.9,
    do_sample=True,
)

outputs = model.generate(**inputs, generation_config=gen_config)

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

print(generated.strip())