Training Data:  https://docs.google.com/spreadsheets/d/1yLDum7yWr3IH0KivluCBEvqHGlfvFW_S/edit?usp=sharing&ouid=107676686611527271344&rtpof=true&sd=true

Test Data: https://docs.google.com/spreadsheets/d/1lCvTufEhGgtDJp6b9oYyFXpCZqWPirSX/edit?usp=sharing&ouid=107676686611527271344&rtpof=true&sd=true

🌿 ArvyaX Machine Learning Internship Assignment

Team ArvyaX · RevoltronX

Theme: From Understanding Humans → To Guiding Them

At ArvyaX, we are building AI systems that go beyond prediction.

We aim to create intelligence that can:

• understand human emotional state

• reason under imperfect and noisy signals

• decide meaningful next actions

• guide users toward better mental states

After immersive sessions (forest, ocean, rain, mountain, café), users write short reflections.

These reflections are:

• messy

• short or vague

• sometimes contradictory

We also collect lightweight contextual signals:

• sleep

• stress

• energy

• time of day

• previous mood

⚠️ Important : This is NOT a standard classification problem.

Real-world systems must handle:

• noisy text

• missing data

• conflicting signals

• imperfect labels

Your goal is to build a system that can understand → decide → guide.

 Objective

Build a system that takes user input and produces:

1. Emotional Understanding

• predicted_state

• predicted_intensity (1–5)

2. Decision Layer (Core)

Your system must decide:

➤ What should the user do?

(e.g., breathing, journaling, deep work, rest)

➤ When should they do it?

• now

• within_15_min

• later_today

• tonight

• tomorrow_morning

3. Uncertainty Awareness

For each prediction:

• confidence (0–1)

• uncertain_flag (0 or 1)

👉 A strong system knows when it is unsure.

4. (Optional Bonus) Supportive Message

Generate a short human-like response explaining the recommendation.

Example:

“You seem slightly restless right now. Let’s slow things down. Try a short breathing exercise before planning your day.”

Dataset

You are provided with a dataset containing: id, journal_text, ambience_type, duration_min, sleep_hours, energy_level, stress_level, time_of_day, previous_day_mood, face_emotion_hint, reflection_quality, emotional_state, intensity

Tasks

Part 1 — Emotional State Prediction

Predict:

• emotional_state

Part 2 — Intensity Prediction

Predict:

• intensity

Explain whether you treat this as:

• classification

• regression

Part 3 — Decision Engine (What + When)

Design logic to decide:

What to do (to make him ready for user intention deep work, sleep etc)

Examples: box_breathing, ournaling. grounding, deep_work, yoga. sound therapy, light_planning, rest, movement, pause

When to do it

Options: now, within_15_min, later_today, tonight, tomorrow_morning

Your system should use: predicted state, intensity, stress, energy, time of day

Part 4 — Uncertainty Modeling

Provide:

• confidence score

• uncertain flag

Part 5 — Feature Understanding

Explain:

• what features mattered most

• text vs metadata importance

Part 6 — Ablation Study

Compare:

• text-only model

• text + metadata model

Part 7 — Error Analysis (Very Important)

Analyze at least 10 failure cases.

Explain:

• what went wrong

• why the model failed

• how to improve

Focus on:

• ambiguous text

• conflicting signals

• short inputs

• noisy labels

Part 8 — Edge / Offline Thinking

Explain how your system would run:

• on mobile

• on-device

Discuss:

• model size

• latency

• tradeoffs

Part 9 — Robustness (Small but Important)

Explain how your system handles:

• very short text (“ok”, “fine”)

• missing values

• contradictory inputs

Constraints

Allowed

• scikit-learn

• XGBoost

• PyTorch

• TensorFlow

• local lightweight models

Not Allowed

• OpenAI / Gemini / Claude APIs

• any hosted LLM

* Your solution must run locally.

Deliverables

1. Code

End-to-end pipeline

2. predictions.csv

• id

• predicted_state

• predicted_intensity

• confidence

• uncertain_flag

• what_to_do

• when_to_do

3. README.md

Include:

• setup instructions

• approach

• feature engineering

• model choice

• how to run

4. ERROR_ANALYSIS.md

Include:

• 10 failure cases

• insights

5. EDGE_PLAN.md

Explain:

• deployment approach

• optimizations

Evaluation Criteria

ML + reasoning - 20%

Decision logic (what + when) - 20%

Uncertainty handling- 15%

Error analysis depth - 15%

Feature understanding- 10%

Code quality- 10%

Edge thinking- 10%

What We Care About

We are looking for:

• real-world thinking

• ability to handle messy data

• reasoning under uncertainty

• meaningful decision-making

• product-oriented mindset

Not just accuracy.

 Bonus (Optional)

• supportive conversational message

• small local API (Flask/FastAPI)

• simple UI demo

• lightweight conversational model (SLM)

• label noise handling

⚠️ Final Note

During the interview, you will be asked to:

• explain your model

• justify decisions

• walk through failure cases

 Philosophy

AI should not just understand humans. It should help them move toward a better state.

Dream > Innovate > Create
— Team ArvyaX