gemma-4-e2b-spatial-lora

Gemma 4 E2B fine-tuned for humanitarian spatial tool calling
Part of GemmaTerrain — Multimodal GeoAI for the Unconnected

Overview

This is a LoRA adapter for google/gemma-4-e2b-it, fine-tuned to output structured JSON tool calls for humanitarian spatial queries. It runs offline on ARM edge hardware (Raspberry Pi 5, Steam Deck) as part of the GemmaTerrain system — no internet, no cloud, no GPU required.

Given a natural language query in English, Bangla, Spanish, or Indonesian, the model selects exactly one spatial tool and outputs its arguments as JSON.

Example:

Input	Output
`"Find the nearest hospital to Camp 6"`	`{"name": "find_nearest_poi_with_route", "arguments": {"poi_type": "hospital", "lat": 21.2, "lon": 92.16}}`
`"15 minute walking radius from Condado"`	`{"name": "generate_isochrone", "arguments": {"lat": 18.46, "lon": -66.07, "max_minutes": 15}}`
`"Camp 6 er kache hospital kothay?"`	`{"name": "find_nearest_poi_with_route", "arguments": {"poi_type": "hospital", "lat": 21.2, "lon": 92.16}}`
`"Farmacias dentro de 1km de Ocean Park"`	`{"name": "list_pois", "arguments": {"poi_type": "pharmacy", "lat": 18.46, "lon": -66.05, "radius_m": 1000}}`

Spatial Tools

The model is trained to call exactly one of these 6 tools:

Tool	Use case
`find_nearest_poi_with_route`	"Nearest hospital to X" — returns closest POI + walking route
`list_pois`	"Clinics within 2km of X" — returns all POIs in radius
`calculate_route`	"Walk from A to B" — returns distance + time
`generate_isochrone`	"15 min walking area from X" — returns reachable boundary
`find_along_route`	"Pharmacies along the way from A to B"
`geocode_place`	"Where is X?" — resolves place name to coordinates

Supported POI types: hospital, clinic, doctors, pharmacy, police, fire_station, shelter, school, university, bank, atm, supermarket, marketplace, drinking_water, water_point, fuel, bus_station, place_of_worship

Training Details

Parameter	Value
Base model	`google/gemma-4-e2b-it`
Method	QLoRA (4-bit quantization)
LoRA rank	16
LoRA alpha	32
Target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`
Max sequence length	512 tokens
Training examples	28 unique × augmented to 5,000
Epochs	3
Batch size	2 (effective: 8 with grad accumulation)
Learning rate	2e-4
Precision	fp16
Framework	Unsloth + TRL SFTTrainer
Hardware	Kaggle T4 (15GB)
Training time	~45 minutes

Dataset

28 hand-crafted humanitarian spatial query → tool call pairs across 3 real-world locations and 4 languages:

Location	Context
Cox's Bazar, Bangladesh	Rohingya refugee camps (Camp 3, 6, 8W, 9, 12, 15)
San Juan, Puerto Rico	Post-hurricane disaster response (Condado, Santurce, Miramar)
Jakarta, Indonesia	Urban humanitarian operations (Menteng, Gelora, Gambir, Kemang)

Languages: English · Bangla (transliterated) · Spanish · Indonesian

Query types: nearest POI, radius search, route calculation, isochrone, along-route search

System Prompt

You are GemmaTerrain, a humanitarian spatial assistant running offline on edge hardware.
Select exactly ONE tool. Output only the tool call JSON — no explanation.

Valid poi_type values: hospital, clinic, doctors, pharmacy, police, fire_station,
shelter, school, university, bank, atm, supermarket, marketplace, drinking_water,
water_point, fuel, bus_station, place_of_worship

Usage

With Unsloth (recommended)

from unsloth import FastModel

model, tokenizer = FastModel.from_pretrained(
    model_name="tasfuuu19/gemma-4-e2b-spatial-lora",
    max_seq_length=512,
    load_in_4bit=True,
)

With Transformers

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("google/gemma-4-e2b-it", load_in_4bit=True)
model = PeftModel.from_pretrained(base, "tasfuuu19/gemma-4-e2b-spatial-lora")
tokenizer = AutoTokenizer.from_pretrained("tasfuuu19/gemma-4-e2b-spatial-lora")

Inference example

messages = [
    {"role": "system", "content": "You are Meridian, a humanitarian spatial assistant..."},
    {"role": "user", "content": "Find the nearest hospital to Camp 6"},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=128, temperature=0.1)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
# {"name": "find_nearest_poi_with_route", "arguments": {"poi_type": "hospital", "lat": 21.2, "lon": 92.16}}

GemmaTerrain System

This adapter is the E2B component of the GemmaTerrain dual-model routing system:

Query → Geocode Layer → Cactus Router → Gemma 4 E2B (this model) → Spatial Tool → Result
                                      ↘ Gemma 4 E4B (complex/multimodal queries)

The router sends queries to E2B for simple spatial lookups (nearest, list, route, isochrone) and to E4B for multimodal or complex multi-hop reasoning. Battery level is also a routing signal — below 20%, all queries go to E2B.

The spatial backend uses NetworKit (Dijkstra routing) + DuckDB (spatial queries) on OpenStreetMap data, running fully offline.

Limitations

Training set is small (28 unique examples) — generalization relies heavily on the base model
Coordinates in training data are fixed to 3 locations; novel locations require geocoding pre-processing
Walking speed hardcoded at 5 km/h (83.33 m/min)
Not suitable for driving or cycling routing

License

Apache 2.0 — same as the base Gemma 4 model.

tasfuuu19/gemma-4-e2b-spatial-lora