voyage-student-1-5b-dayfill

voyage-student-1-5b-dayfill

Distilled 1.5B travel-planning student for the Local LLM AI Agent for Travel Planning desktop app.

Fine-tuned from Qwen/Qwen2.5-1.5B-Instruct with supervised fine-tuning (SFT) on Trip.Planner-style tool trajectories (route proposal → plan_trip → day-by-day itinerary). Optimized for day-fill itinerary structure and local Ollama serving.

Files

Quick start (Ollama)

# From this repo folder (after downloading the GGUF + Modelfile):
ollama create voyage-student-1-5b-dayfill -f Modelfile
ollama run voyage-student-1-5b-dayfill

Or with Hugging Face Hub:

huggingface-cli download ChristopherLi/voyage-student-1-5b-dayfill \
  student-1-5b-dayfill.q4_k_m.gguf Modelfile \
  --local-dir ./voyage-student-1-5b-dayfill
cd voyage-student-1-5b-dayfill
ollama create voyage-student-1-5b-dayfill -f Modelfile

Point the travel agent at it (default):

# .env
OLLAMA_MODEL=voyage-student-1-5b-dayfill
STUDENT_MODE=true
OLLAMA_NUM_CTX=16384

python -m travel_agent

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "ChristopherLi/voyage-student-1-5b-dayfill"
tok = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto", device_map="auto", trust_remote_code=True)

Intended use

• Local travel itinerary drafting with tool calls against Trip.com scrapes (via the companion app).
• Not a booking engine; prices and URLs must come from live tools, not model invention.

Training

• Base: Qwen2.5-1.5B-Instruct
• Method: SFT (TRL) on distilled teacher rollouts with day-fill emphasis
• Quantization shipped here: Q4_K_M GGUF for Ollama

Citation / attribution

• Base model: Qwen2.5 by Alibaba
• App: An-AI-Agent-for-travel-planning