rakhasetiawan/flint-granite-survival

FLINT — Field Language Intelligence - v.03 (LoRA Adapter)

⚠️ Portfolio project. Built as a learning exercise to demonstrate a full CPT + LoRA fine-tuning pipeline. Not intended for production use.

Mission

FLINT was initiated to explore the feasibility of an offline-first AI carrying a survival domain knowledge base — intended for search and rescue personnel, field operators, and anyone working in environments without internet connectivity. The goal is a model that can be deployed on-device and queried without any network dependency, providing survival-critical information when it matters most.

This is the LoRA adapter for FLINT, a survival-domain language model continued pre-trained on Wikipedia corpus. Load this adapter on top of ibm-granite/granite-4.0-micro to use FLINT.

Training Details

Model Architecture (Verified)

PeftModelForCausalLM
  └── LoraModel
        └── GraniteMoeHybridForCausalLM
              ├── Attention layers
              │     └── q_proj:        lora.Linear (r=8)
              │     └── k_proj:        lora.Linear (r=8)
              │     └── v_proj:        lora.Linear (r=8)
              │     └── o_proj:        lora.Linear (r=8)
              └── MLP layers
                    └── input_linear:  lora.Linear (r=8)
                    └── output_linear: lora.Linear (r=8)

How to Use

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

# Load base model
tokenizer = AutoTokenizer.from_pretrained("ibm-granite/granite-4.0-micro")
model     = AutoModelForCausalLM.from_pretrained(
    "ibm-granite/granite-4.0-micro",
    device_map = "cpu",  # change to "cuda" if GPU available
)

# Load FLINT adapter on top
model = PeftModel.from_pretrained(model, "rakhasetiawan/flint-granite-survival")
model.eval()

# Prompt it like a Wikipedia article opening
prompt = "Wilderness survival techniques include"
inputs = tokenizer(prompt, return_tensors="pt")

outputs = model.generate(
    **inputs,
    max_new_tokens = 200,
    temperature    = 0.7,
    do_sample      = True,
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Limitations

• —CPT only — not instruction-tuned, no conversational ability
• —Short training run (~250–300 steps), knowledge depth is limited
• —Prototype quality — outputs may be rough or repetitive
• —Next step would be SFT on survival-domain Q&A pairs

Training Loss

Loss dropped from 5.04 → ~3.05 over 250 steps, with the curve stabilizing around step 100. The smooth convergence indicates stable training with no instability or gradient issues.

Prompt Example

"what do i do to prepare for a natural disaster? give me a checklist of items
to prepare and steps to take"

Response

what do i do to prepare for a natural disaster? give me a checklist of items to prepare and steps to take.

1. 1.Assess Your Risks: Before you start preparing, determine which natural disasters are most likely to occur in your area. This could include hurricanes, earthquakes, floods, wildfires, or tornadoes. Knowing your risks will help you understand what preparations and protective measures you need to take.

1. 1.Create a Disaster Preparedness Plan: Draft a plan that includes communication strategies in case family members are separated, evacuation routes, and safe meeting points. Make sure everyone in your household knows what to do in an emergency.

1. 1.Build an Emergency Kit: Assemble a kit that includes essentials you might need for at least 72 hours. Here's a checklist of items to include:

Eval Notes

Note: Quantified figures (1 gallon/day, 72-hour window) align with FEMA standard preparedness guidelines — not hallucinated values. Practitioner-level depth (disaster-specific kits, document preservation, post-disaster protocols) is the target for v0.4 with field manual corpus.

FLINT — Field Language Intelligence · Built by [rakhasetiawan](https://huggingface.co/rakhasetiawan)