nxvoy-labs/shasa-v0.3

Shasa v0.3 — NxVoy's Travel-Native AI

The world's first travel-native foundation model

🌍 Itinerary Generation · 💬 Travel Chat · 🎯 Intent Classification · 📝 Smart Clarification

Model Overview

Shasa is NxVoy's proprietary travel AI model, fine-tuned via QLoRA on curated travel data covering 6 specialized capabilities. Unlike API wrappers that call generic LLMs, Shasa is a domain-specific model with proprietary weights optimized for travel planning.

Property	Value
Base Model	Qwen/Qwen2.5-3B-Instruct
Method	QLoRA (4-bit NF4 quantization)
LoRA Rank	64
LoRA Alpha	128
Target Modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Trainable Params	~80M / 3.1B (2.6%)
Training Data	2,550 curated travel examples
Training Time	3h 08m on NVIDIA L40S (48GB)
Epochs	5
Learning Rate	2e-4 (cosine schedule with warmup)
Batch Size	2 (gradient accumulation: 4)

Training Metrics

Final Training Loss:      0.1575
Final Eval Loss:          0.2247
Token Accuracy (Train):   95.06%
Token Accuracy (Eval):    93.54%
Total Steps:              800
Training Time:            3h 08m 23s
GPU:                      NVIDIA L40S (48GB)

Loss Progression

Epoch	Train Loss	Eval Loss	Token Accuracy
1	0.3521	0.2891	88.2%
2	0.2156	0.2445	92.1%
3	0.1832	0.2301	93.8%
4	0.1648	0.2259	94.7%
5	0.1575	0.2247	95.1%

6 Capabilities

Shasa handles 6 specialized travel tasks, each with a dedicated system prompt:

1. 🗺️ Itinerary Generation

Generates complete trip itineraries as structured JSON with flights, hotels, day plans, activities, meals, and budget breakdowns.

2. 📄 Itinerary Formatting

Converts itinerary JSON into beautiful, readable markdown with emojis and tables.

3. 🎯 Intent Classification

Classifies user messages into 7 intents: new_trip, modify_itinerary, ask_question, select_option, greeting, farewell, off_topic.

4. 💬 Conversational Chat

Multi-turn travel planning conversations with destination recommendations and travel tips.

5. ❓ Smart Clarification

Generates clarifying questions to gather missing trip details (dates, budget, preferences).

6. 🏷️ Thread Naming

Generates short, descriptive thread names for travel conversations (max 50 chars).

Usage

With PEFT + Transformers

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

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

# Generate itinerary
messages = [
    {"role": "system", "content": "You are Shasa, NxVoy's expert travel AI. Generate a complete trip itinerary as structured JSON."},
    {"role": "user", "content": "Plan a 5-day trip to Tokyo for 2 adults, budget $3000 USD"}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=4096, temperature=0.3)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

With vLLM (Production)

from vllm import LLM, SamplingParams

llm = LLM(
    model="Qwen/Qwen2.5-3B-Instruct",
    enable_lora=True,
    max_lora_rank=64,
)

output = llm.generate(
    prompts,
    SamplingParams(temperature=0.3, max_tokens=4096),
    lora_request=LoRARequest("shasa", 1, "nxvoy-labs/shasa-v0.3"),
)

Training Data

The model was trained on 2,550 curated examples across all 6 capabilities:

Capability	Examples	Source
Itinerary Generation	~400	Gemini 2.5 Flash distillation
Intent Classification	~400	Synthetic + manual
Conversational Chat	~400	Travel conversations
Smart Clarification	~400	Question generation
Destination Knowledge	~400	Travel Q&A
Thread Naming	~350	Short text generation
Itinerary Formatting	~200	JSON → Markdown

All data is in chat format with system/user/assistant messages following the Qwen chat template.

Architecture

NxVoy Production Stack:
┌──────────────────────────────┐
│  NxVoy API (FastAPI)         │
├──────────────────────────────┤
│  NxVoy-FM Router             │
│  ├── Shadow / AB / Primary   │
│  ├── Shasa → Gemini fallback │
│  └── Schema validation       │
├──────────────────────────────┤
│  vLLM + LoRA hot-swap        │
├──────────────────────────────┤
│  GPU (L4/A10G, 4-bit GPTQ)  │
└──────────────────────────────┘

Version History

Version	Base	Data	Key Change
v0.1	Qwen 2.5 3B	850	Proof of concept
v0.2	Qwen 2.5 3B	2,550	All 6 capabilities
v0.3	Qwen 2.5 3B	2,550	Fixed training config, L40S GPU
v0.4 (planned)	Qwen 2.5 7B	50K	Scale up base + data

Limitations

• Small training set (2,550 examples) — may generate generic responses
• 3B parameter model — can struggle with very complex itinerary JSON
• May hallucinate hotel/restaurant names (no grounding database yet)
• English-only training data
• Not yet evaluated against production benchmarks

License

Apache 2.0 — free for commercial use.

Citation

@misc{shasa2026,
  title={Shasa: A Travel-Native Foundation Model},
  author={NxVoy Labs},
  year={2026},
  url={https://huggingface.co/nxvoy-labs/shasa-v0.3}
}

---

Built by NxVoy — Intelligent Travel Planning