Spaces:
Sleeping
Sleeping
Upload 2 files
Browse files- app.py +199 -0
- requirements.txt +8 -0
app.py
ADDED
|
@@ -0,0 +1,199 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
VisualStep β ADHD Classroom Instruction Decomposer
|
| 3 |
+
Gradio demo app for Hugging Face Spaces
|
| 4 |
+
"""
|
| 5 |
+
|
| 6 |
+
import json
|
| 7 |
+
import gradio as gr
|
| 8 |
+
import torch
|
| 9 |
+
from unsloth import FastModel
|
| 10 |
+
import spaces
|
| 11 |
+
|
| 12 |
+
|
| 13 |
+
# ββ Model loading βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 14 |
+
MODEL_ID = "lsadouk1111/VisualStep"
|
| 15 |
+
|
| 16 |
+
print("Loading VisualStep model...")
|
| 17 |
+
model, tokenizer = FastModel.from_pretrained(
|
| 18 |
+
model_name=MODEL_ID,
|
| 19 |
+
max_seq_length=512,
|
| 20 |
+
load_in_4bit=True,
|
| 21 |
+
dtype=None,
|
| 22 |
+
)
|
| 23 |
+
FastModel.for_inference(model)
|
| 24 |
+
model.eval()
|
| 25 |
+
print("Model ready.")
|
| 26 |
+
|
| 27 |
+
# ββ System prompt βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 28 |
+
SYSTEM_PROMPT = """You are an expert in ADHD classroom accommodations for primary school children aged 5-10.
|
| 29 |
+
Transform the teacher's spoken classroom instruction into an ADHD-adapted visual step card.
|
| 30 |
+
|
| 31 |
+
Rules:
|
| 32 |
+
- Maximum 5 steps
|
| 33 |
+
- Each step: one concrete physical action only
|
| 34 |
+
- Each action: maximum 6 words
|
| 35 |
+
- Start each step with an imperative verb
|
| 36 |
+
- Include a pictogram keyword in the icon field
|
| 37 |
+
- Output JSON only, no other text
|
| 38 |
+
|
| 39 |
+
Output format:
|
| 40 |
+
{"steps":[{"id":1,"action":"Open your book","detail":"page 23","icon":"book","check":true}],"n_steps":1,"support_level":"medium"}
|
| 41 |
+
|
| 42 |
+
support_level: "light" (1-2 steps), "medium" (3-4 steps), "high" (5 steps)"""
|
| 43 |
+
|
| 44 |
+
# ββ Icon to emoji ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 45 |
+
EMOJI = {
|
| 46 |
+
"book": "π", "page": "π", "pencil": "βοΈ", "write": "βοΈ",
|
| 47 |
+
"read": "ποΈ", "eye": "ποΈ", "hand": "β", "finger": "π",
|
| 48 |
+
"talk": "π£οΈ", "listen": "π", "think": "π§ ", "look": "ποΈ",
|
| 49 |
+
"draw": "π¨", "paper": "π", "notebook": "π", "board": "π",
|
| 50 |
+
"number": "π’", "count": "π’", "answer": "β
", "sort": "ποΈ",
|
| 51 |
+
"match": "π", "circle": "β", "plant": "π±", "seed": "π°",
|
| 52 |
+
"water": "π§", "animal": "πΎ", "picture": "πΌοΈ", "partner": "π₯",
|
| 53 |
+
"find": "π", "search": "π", "observe": "π", "measure": "π",
|
| 54 |
+
"record": "π", "label": "π·οΈ", "colour": "ποΈ", "color": "ποΈ",
|
| 55 |
+
"cut": "βοΈ", "touch": "π", "quiet": "π€«", "ready": "β
",
|
| 56 |
+
"open": "π", "put": "π", "place": "π", "sit": "πͺ",
|
| 57 |
+
"science": "π¬", "maths": "β", "english": "π",
|
| 58 |
+
}
|
| 59 |
+
|
| 60 |
+
def get_emoji(icon):
|
| 61 |
+
if not icon:
|
| 62 |
+
return "π"
|
| 63 |
+
k = str(icon).lower()
|
| 64 |
+
for key, em in EMOJI.items():
|
| 65 |
+
if key in k:
|
| 66 |
+
return em
|
| 67 |
+
return "π"
|
| 68 |
+
|
| 69 |
+
# ββ Generation ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 70 |
+
@spaces.GPU
|
| 71 |
+
def generate_step_card(instruction, year, subject):
|
| 72 |
+
if not instruction.strip():
|
| 73 |
+
return "Please enter a classroom instruction."
|
| 74 |
+
|
| 75 |
+
messages = [
|
| 76 |
+
{"role": "system", "content": SYSTEM_PROMPT},
|
| 77 |
+
{"role": "user", "content": f"Year: {year} | Subject: {subject} | Instruction: {instruction}"},
|
| 78 |
+
]
|
| 79 |
+
prompt = tokenizer.apply_chat_template(
|
| 80 |
+
messages, tokenize=False, add_generation_prompt=True
|
| 81 |
+
)
|
| 82 |
+
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
|
| 83 |
+
|
| 84 |
+
with torch.no_grad():
|
| 85 |
+
outputs = model.generate(
|
| 86 |
+
**inputs,
|
| 87 |
+
max_new_tokens=256,
|
| 88 |
+
do_sample=False,
|
| 89 |
+
pad_token_id=tokenizer.eos_token_id,
|
| 90 |
+
)
|
| 91 |
+
generated = outputs[0][inputs["input_ids"].shape[1]:]
|
| 92 |
+
raw = tokenizer.decode(generated, skip_special_tokens=True).strip()
|
| 93 |
+
raw = raw.replace("```json", "").replace("```", "").strip()
|
| 94 |
+
|
| 95 |
+
try:
|
| 96 |
+
data = json.loads(raw)
|
| 97 |
+
steps = data.get("steps", [])
|
| 98 |
+
support = data.get("support_level", "medium")
|
| 99 |
+
n = data.get("n_steps", len(steps))
|
| 100 |
+
|
| 101 |
+
# Build visual card
|
| 102 |
+
lines = ["## π WHAT TO DO NOW\n"]
|
| 103 |
+
for s in steps:
|
| 104 |
+
emoji = get_emoji(s.get("icon", ""))
|
| 105 |
+
action = s.get("action", "")
|
| 106 |
+
detail = s.get("detail", "")
|
| 107 |
+
line = f"β **{s.get('id', '')}.** {emoji} {action}"
|
| 108 |
+
if detail:
|
| 109 |
+
line += f" β *{detail}*"
|
| 110 |
+
lines.append(line)
|
| 111 |
+
|
| 112 |
+
lines.append(f"\n---")
|
| 113 |
+
lines.append(f"*Support level: {support} Β· {n} step(s)*")
|
| 114 |
+
|
| 115 |
+
return "\n\n".join(lines)
|
| 116 |
+
|
| 117 |
+
except json.JSONDecodeError:
|
| 118 |
+
return f"β οΈ Model output could not be parsed as JSON.\n\nRaw output:\n```\n{raw}\n```"
|
| 119 |
+
|
| 120 |
+
# ββ Gradio interface ββββββββββββββββββββββββββββββββββββββββββββββββββββββββοΏ½οΏ½οΏ½ββ
|
| 121 |
+
EXAMPLES = [
|
| 122 |
+
["Ok everyone, open your books to page 23, read the paragraph quietly, then answer questions 1 to 3 in your notebook and put your hand up when you're done.", "2", "english"],
|
| 123 |
+
["Right, so, um, can we all take our science books and turn to page 10? And then, after that, I want you to look at the picture and think about what you see.", "1", "science"],
|
| 124 |
+
["Everyone stop what you're doing, pack away your things, and line up quietly at the door please.", "3", "routine"],
|
| 125 |
+
["I want you to sort these animals into mammals and not mammals using the sorting hoops, then write one sentence explaining how you decided.", "1", "science"],
|
| 126 |
+
["OK so, find your maths book, open to page 15, do the first three problems and check your answers with your partner.", "2", "maths"],
|
| 127 |
+
]
|
| 128 |
+
|
| 129 |
+
with gr.Blocks(
|
| 130 |
+
title="VisualStep β ADHD Instruction Decomposer",
|
| 131 |
+
theme=gr.themes.Soft(primary_hue="blue"),
|
| 132 |
+
css="""
|
| 133 |
+
.card-output { font-size: 1.1em; line-height: 1.8; }
|
| 134 |
+
h1 { color: #1F4E79; }
|
| 135 |
+
.subtitle { color: #666; font-size: 0.95em; margin-top: -10px; }
|
| 136 |
+
"""
|
| 137 |
+
) as demo:
|
| 138 |
+
|
| 139 |
+
gr.Markdown("""
|
| 140 |
+
# VisualStep
|
| 141 |
+
### ADHD-Adapted Visual Instruction Decomposition for Primary School Classrooms
|
| 142 |
+
|
| 143 |
+
*Fine-tuned Phi-3 Mini (3.8B) Β· Lamyaa Sadouk Β· EMSI Casablanca*
|
| 144 |
+
|
| 145 |
+
---
|
| 146 |
+
|
| 147 |
+
Enter a classroom instruction as a teacher would say it out loud.
|
| 148 |
+
VisualStep will decompose it into a structured, ADHD-adapted visual step card.
|
| 149 |
+
""")
|
| 150 |
+
|
| 151 |
+
with gr.Row():
|
| 152 |
+
with gr.Column(scale=1):
|
| 153 |
+
instruction = gr.Textbox(
|
| 154 |
+
label="Teacher's spoken instruction",
|
| 155 |
+
placeholder="e.g. Ok everyone, open your books to page 23, read the paragraph quietly, then answer questions 1 to 3...",
|
| 156 |
+
lines=4,
|
| 157 |
+
)
|
| 158 |
+
with gr.Row():
|
| 159 |
+
year = gr.Dropdown(
|
| 160 |
+
choices=["1", "2", "3", "4", "5", "6"],
|
| 161 |
+
value="2",
|
| 162 |
+
label="Year group",
|
| 163 |
+
)
|
| 164 |
+
subject = gr.Dropdown(
|
| 165 |
+
choices=["english", "maths", "science", "routine", "transition"],
|
| 166 |
+
value="english",
|
| 167 |
+
label="Subject",
|
| 168 |
+
)
|
| 169 |
+
btn = gr.Button("Generate Step Card", variant="primary", size="lg")
|
| 170 |
+
|
| 171 |
+
with gr.Column(scale=1):
|
| 172 |
+
output = gr.Markdown(
|
| 173 |
+
label="ADHD-adapted step card",
|
| 174 |
+
elem_classes=["card-output"],
|
| 175 |
+
value="*Your step card will appear here.*"
|
| 176 |
+
)
|
| 177 |
+
|
| 178 |
+
btn.click(
|
| 179 |
+
fn=generate_step_card,
|
| 180 |
+
inputs=[instruction, year, subject],
|
| 181 |
+
outputs=output,
|
| 182 |
+
)
|
| 183 |
+
|
| 184 |
+
gr.Examples(
|
| 185 |
+
examples=EXAMPLES,
|
| 186 |
+
inputs=[instruction, year, subject],
|
| 187 |
+
label="Try these examples",
|
| 188 |
+
)
|
| 189 |
+
|
| 190 |
+
gr.Markdown("""
|
| 191 |
+
---
|
| 192 |
+
**About VisualStep:**
|
| 193 |
+
This demo accompanies the paper *"VisualStep: A Fine-Tuned Small Language Model for ADHD-Adapted
|
| 194 |
+
Visual Instruction Decomposition in Primary School Classrooms"*.
|
| 195 |
+
The model was fine-tuned on VisualStep-2K, a dataset of 2,000 spoken classroom instructionβstep card pairs.
|
| 196 |
+
All inference runs locally β no data is sent to external servers.
|
| 197 |
+
""")
|
| 198 |
+
|
| 199 |
+
demo.launch()
|
requirements.txt
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
spaces
|
| 2 |
+
unsloth
|
| 3 |
+
torch
|
| 4 |
+
gradio
|
| 5 |
+
transformers
|
| 6 |
+
peft
|
| 7 |
+
bitsandbytes
|
| 8 |
+
accelerate
|