agurusantosh/tripmind-ft-lora

tripmind-ft

Fine-tuned Llama 3.1 8B for Indian domestic travel optimization. Given a traveler persona, generates an optimized day-by-day itinerary identifying Price-Pivot Points — transit, accommodation, or activity substitutions that save ≥5% without degrading trip quality.

Part of the TripMind project: a multi-agent AI travel optimizer trained via three distinct approaches (SFT, distillation, curriculum). tripmind-ft is the best-performing variant, trained via standard supervised fine-tuning on 5,000 synthetic pairs generated by GPT-4o-mini.

Model Details

Property	Value
Base model	unsloth/Meta-Llama-3.1-8B
Training method	QLoRA r=8, α=16, dropout=0.05
Training data	4,749 Alpaca-format pairs (Phase 1 synthetic)
Epochs	3
Final train loss	0.266
Hardware	Colab T4 (fp16, seq_len=512)
Format	GGUF Q4_K_M (4.6 GB)

Evaluation Results (92 test cases)

Metric	Score	Target	✓/✗
JSON valid	100%	85%	✓
Savings found	100%	70%	✓
Budget compliance	98.7%	80%	✓
Schema compliance	83.7%	80%	✓
BERTScore F1	0.932	0.70	✓
ROUGE-L	0.436	0.25	✓
Reasoning coherence	0.723	0.65	✓
Grounding accuracy	0.895	0.60	✓
Intent alignment	0.322	0.55	✗
Red-team pass	53.3%	80%	✗

Head-to-head: beats tripmind-distill 78% of the time, tripmind-curriculum 57%.

Usage with Ollama

# Download GGUF from this repo
ollama create tripmind-ft -f Modelfile.ft

# Run
ollama run tripmind-ft

Prompt format (Alpaca):

### Instruction:
Act as TripMind Optimizer. Given a traveler persona for an Indian domestic trip, produce an optimized day-by-day itinerary that minimizes total cost while respecting the budget tier, trip type, and traveler intents. Identify the primary Price-Pivot Point (transit, accommodation, or activity substitution that saves ≥5%) and explain it clearly.

### Input:
{"starting_city": "Mumbai", "destination_city": "Delhi", "type": "Solo", "size": {"adults": 1, "children": 0}, "intents": ["Adventure"], "budget": "Shoestring", "duration_days": 5, "duration_nights": 4}

### Response:

Limitations

• Trained on Indian domestic travel only (20 cities). Not designed for international travel.
• Red-team robustness is below target (53.3% vs 80% goal) — the model can be prompted to bypass budget constraints.
• Intent alignment is below target (32.2% vs 55%) — cost optimization is prioritized over activity personalization.
• Inference on CPU takes 30–120 seconds per query (use GPU for production).