likas-gemma4-e2b-lora

LIKAS — Gemma 4 E2B-IT (Disaster Companion for the Philippines)

LoRA fine-tune of google/gemma-4-e2b-it that turns the base model into LIKAS, an offline disaster-response assistant for the Philippines. The model emits a strict JSON envelope for every turn — either a tool call or a user-facing reply — so it can be safely run on-device inside a React Native app via llama.rn.

This is the model that powers LIKAS, an offline-first mobile app for earthquake / typhoon / volcanic-eruption preparedness, evacuation routing, and POI search across Metro Manila.

⚠️ This is a fine-tune for a specific app, not a general-purpose chatbot. It only emits structured JSON and is grounded in NDRRMC / PAGASA / PHIVOLCS protocols. Use it inside the dispatch loop described below or it will produce JSON your client must parse.

Repository variants

This repository hosts the LoRA adapter. The full pipeline produces three artifacts:

Intended use

In scope:

• Disaster preparedness, response, and recovery Q&A for earthquake, typhoon, and volcanic eruption in the Philippines
• Quoting NDRRMC / PAGASA / PHIVOLCS protocols verbatim via a get_protocol tool
• Ranking the nearest evacuation centers via a route_to_nearest_evacuation tool (returns a polyline the UI draws on a MapLibre map)
• Finding nearby hospitals, schools, gymnasiums, multi-purpose halls, covered courts via find_nearby
• Personalizing replies from on-device user profile data via get_user_profile
• Multilingual replies in English, Filipino, and Taglish (replies in the user's language)
• Polite refusal of off-topic queries (entertainment, opinions, general knowledge)

Out of scope:

• Any non-disaster topic — the model refuses by design (Rule 7)
• Calling, SMS, or any real-world action — the model never claims to do these (Rule 9)
• Medical diagnosis — the model can reference profile-stored conditions but does not diagnose
• Domains outside the Philippines (NDRRMC/PAGASA/PHIVOLCS guidance is country-specific)
• Use as a free-form chatbot — the output grammar is constrained to a JSON envelope

Output contract

Every turn the model produces exactly one JSON object on a single line. There is no prose outside the JSON. Two shapes:

{"action": "tool", "name": "<tool_name>", "args": {...}}

{"action": "speak", "text": "<reply to user>"}

After the host runs the tool and injects the result, the model decides again — either another tool call or a final speak turn. The host typically caps the tool-dispatch loop at 3 iterations per user query.

Tools

Example exchange (Tagalog)

user:     Bagyo na, kailangan kong lumikas
assistant {"action":"tool","name":"get_protocol","args":{"disaster":"typhoon","phase":"during"}}
tool      {"name":"get_protocol","result":"PAGASA and NDRRMC guidance: Stay indoors..."}
assistant {"action":"speak","text":"Mga puntos mula sa PAGASA: Stay indoors away from windows..."}
user:     Saan ako lilikas?
assistant {"action":"tool","name":"route_to_nearest_evacuation","args":{"profile_aware":true}}
tool      {"name":"route_to_nearest_evacuation","result":"Top evacuation options: 1. Manila High School..."}
assistant {"action":"speak","text":"Pinakamalapit: Manila High School (0.8 km, ~11 min lakad). Naka-plot na sa mapa ang ruta..."}

Training data

Trained on jeypiic/likas-ai-datasets (v4) — 691 conversations stratified across 10 task categories, with assistant-output uniqueness of ~80% on speak turns (compared to 30% on v3, which overfit).

User origins are sampled from 20 Metro Manila barangay centroids; POIs are loaded from OSM extracts (59 hospitals, 75 schools, etc.) inside bbox=(120.85, 14.25, 121.30, 14.90). Distances and walking ETAs are computed via haversine + a 4.2 km/h pace, matching the production routingService exactly.

The runtime tool result strings are 1:1 with what Likas/src/services/aiTools.ts produces, so the model sees the same distribution at inference time as during training.

Training procedure

• Base model: unsloth/gemma-4-E2B-it (4-bit quantized for training)
• Method: PEFT LoRA SFT via Unsloth
• Library: transformers==5.5.0, trl (Unsloth pin), peft, bitsandbytes, unsloth_zoo
• Compute: Kaggle T4 (single GPU)
• Hyperparameters:
  • LoRA r=32, alpha=32, dropout=0
  • Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
  • Optimizer: adamw_8bit, learning rate 2e-4, cosine schedule, warmup ratio 0.03
  • Batch size 2 × grad accumulation 4 (effective batch 8)
  • Max sequence length 4096 (system prompt is long; tool turns add up)
  • 3 epochs (~233 optimizer steps total on 622 train rows)
  • Loss computed over the full formatted sequence (full-sequence SFT)
  • Gradient checkpointing via Unsloth, max_grad_norm=0.3, weight_decay=0.001
  • Seed 3407

Per-step train and validation loss were logged to Weights & Biases. The full notebook is at Likas_Gemma4_E2B_Finetune.ipynb.

How to use

As LoRA on top of base Gemma 4 E2B-IT (Python)

from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="jpcurada/likas-gemma4-e2b-lora",
    max_seq_length=4096,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

SYSTEM_PROMPT = """..."""  # see README of the dataset for the exact prompt
messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user",   "content": "May lindol! Ano gagawin ko?"},
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
out = model.generate(**inputs, max_new_tokens=256, temperature=1.0, top_p=0.95, top_k=64)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

On mobile via llama.rn (production)

Download likas-q4_k_m.gguf from jpcurada/likas-gemma4-e2b-gguf and load via llama.rn. The LIKAS app constrains generation with a GBNF grammar (Likas/src/services/aiGrammar.ts) so the model is forced to emit valid JSON envelopes even at temperature 1.0.

Evaluation

Loss curves are logged to wandb during training. We do not publish a benchmark score because the model is task-specific (Filipino disaster-response tool-calling); no standard benchmark covers this distribution.

Manual evaluation rubric used in development:

1. JSON validity — every generated string parses as a single JSON object
2. Tool selection — get_protocol for safety-critical, route_to_nearest_evacuation for evacuation, find_nearby for POI, get_user_profile for self-queries
3. Rule coverage — speak turns include every rule from the protocol tool result (no silent drops)
4. Language match — Rule 8 — reply in the user's language
5. Refusal handling — Rule 7 — politely redirect off-topic queries
6. Map awareness — when route_state == "route_ok", the speak turn references the polyline drawn on the map

Limitations & risks

• Geographic scope: Trained only on Metro Manila POIs and origins. Behavior outside the pedestrian-graph bbox (120.85, 14.25, 121.30, 14.90) is undefined.
• Authority drift: Protocol texts are a snapshot of NDRRMC / PAGASA / PHIVOLCS guidance at training time. If authorities update guidance, the model will quote the older text until re-fine-tuned. Always defer to live authority broadcasts in an actual emergency.
• Not a substitute for emergency services: This is a preparedness companion, not a replacement for calling 911 (NDRRMC) or 117 (BFP fire emergencies).
• JSON envelope: The model emits a strict JSON envelope. If you call it without a JSON parser / GBNF grammar / system prompt that matches the training distribution, output may not be what you expect.
• No real-world actions: The model never sends SMS, places calls, or takes actions outside its host app. Any tool call shown in this README is mediated by the host application.
• Small model: ~5B params at fp16. Capable for structured tool-use over a narrow domain, not a general assistant.

License

Distributed under the Gemma Terms of Use (inherited from the base model). You must accept Google's Gemma license before using this adapter. Training data is the author's own work (jeypiic/likas-ai-datasets) and is governed by that dataset's license.

Citation

@misc{likas-gemma4-e2b-2026,
  author = {John Paul Curada},
  title  = {LIKAS — Gemma 4 E2B-IT fine-tune for offline disaster response in the Philippines},
  year   = {2026},
  url    = {https://huggingface.co/jpcurada/likas-gemma4-e2b-lora},
}

Acknowledgments

• Google for Gemma 4
• Unsloth for the fast LoRA SFT pipeline
• NDRRMC, PAGASA, PHIVOLCS for the authoritative protocols this model is grounded in
• OpenStreetMap contributors (via Geofabrik Philippines extract) for the POIs and pedestrian graph