Shasa is a purpose-built travel AI model by NxVoy. It's a LoRA fine-tune of Qwen2.5-3B-Instruct, trained on 789 curated travel examples distilled from GPT-4o, Claude 3.5 Sonnet, and Gemini Flash.

Model Details


Base Model	Qwen/Qwen2.5-3B-Instruct
Method	LoRA (r=64, alpha=128)
Training	SFT, 3 epochs, 252 steps
Dataset	nxvoy-labs/shasa-travel-distillation-v1 (789 examples)
Parameters	119M trainable / 3.2B total (3.7%)
Hardware	NVIDIA A10G (HuggingFace Spaces)
Framework	transformers + PEFT + TRL
License	Apache 2.0
Developed by	NxVoy Labs

7 Travel Capabilities

🗺️ Itinerary Generation — Day-by-day structured travel plans with timing, costs, tips
❓ Clarification Dialogue — Smart follow-up questions to refine vague requests
🏙️ Destination Intelligence — Factual descriptions grounded in real data
💰 Budget Optimization — Budget/mid-range/luxury breakdowns with cost estimates
📋 JSON Output — Structured machine-readable itinerary format
🎯 Intent Classification — Detect trip planning vs chat vs booking intent
🧵 Thread Naming — Auto-generate conversation titles

Usage

python

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

# Load base model + LoRA adapter
base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-3B-Instruct",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
model = PeftModel.from_pretrained(base, "nxvoy-labs/shasa-v0.2")
tokenizer = AutoTokenizer.from_pretrained("nxvoy-labs/shasa-v0.2")

# Chat
messages = [
    {"role": "system", "content": "You are Shasa, NxVoy's travel AI assistant. You create detailed, practical travel itineraries."},
    {"role": "user", "content": "Plan a 3-day trip to Tokyo for 2 adults, mid-range budget"}
]

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

with torch.no_grad():
    output = model.generate(
        **inputs,
        max_new_tokens=2048,
        temperature=0.7,
        do_sample=True,
        top_p=0.9
    )

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

Training Details

Dataset

789 examples across 7 capabilities
Multi-teacher distillation: GPT-4o (40%), Claude 3.5 Sonnet (35%), Gemini Flash (25%)
Quality filtered: Only examples scoring ≥0.85 quality score retained
Destinations covered: 50+ cities across 6 continents

Hyperparameters

Parameter	Value
LoRA rank (r)	64
LoRA alpha	128
LoRA dropout	0.05
Learning rate	2e-4
Epochs	3
Batch size (effective)	8
Max sequence length	4096
Optimizer	AdamW
Scheduler	Cosine
Warmup ratio	3%
Precision	bf16

Target Modules

q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

Intended Use

Shasa v0.2 is designed for:

Travel itinerary generation — Creating day-by-day trip plans
Travel Q&A — Answering destination questions
Trip planning chatbot — Multi-turn conversation for trip refinement
Structured output — JSON itinerary generation for app integration

Limitations

Trained on 789 examples — production deployment should use v0.3+ with 10K+ examples
Best for English travel queries
Should be paired with retrieval (e.g., Firestore destination data) for factual accuracy
Not intended for booking or financial transactions

About NxVoy

NxVoy builds AI-powered travel planning technology. Shasa is our foundation model for travel intelligence, designed to compete with the world's best travel AI systems.

Citation

bibtex

@misc{shasa-v0.2-2026,
  title={Shasa v0.2: A Travel-Specialized Language Model},
  author={NxVoy Labs},
  year={2026},
  publisher={HuggingFace},
  url={https://huggingface.co/nxvoy-labs/shasa-v0.2}
}