sdt-gm4-hesabu-flare-v0.01

sdt-gm4-hesabu-flare-v0.01

Swahili math-assistant LoRA adapter focused on primary-school arithmetic and word problems.

Model details

• Model name: sdt-gm4-hesabu-flare-v0.01
• Base model: unsloth/gemma-3-4b-it-unsloth-bnb-4bit
• Fine-tuning approach: QLoRA / LoRA adapter fine-tuning
• Primary language: Swahili
• Target domain: School-level mathematics tutoring

Training data

This model was trained with supervised chat-style examples from:

• SFT train/validation splits (train.jsonl, val.jsonl)
• Arithmetic booster set (math_booster.jsonl)

Associated dataset repo:

• Japhari/sdt-gm4-hesabu-flare-v0.01-dataset

Intended use

• Swahili math Q&A assistance
• Educational demos and experimentation
• Baseline model for further instruction tuning

Limitations

• May still hallucinate on complex multi-step reasoning
• Performance degrades on prompts missing key quantities
• Not suitable as sole source for high-stakes grading/financial decisions

Safety notes

For production or assessment workflows, use deterministic validation/guardrails for arithmetic final answers.

Inference example (Transformers + PEFT)

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = "unsloth/gemma-3-4b-it-unsloth-bnb-4bit"
adapter_repo = "Japhari/sdt-gm4-hesabu-flare-v0.01"

tokenizer = AutoTokenizer.from_pretrained(adapter_repo)
model = AutoModelForCausalLM.from_pretrained(base_model, device_map="auto")
model = PeftModel.from_pretrained(model, adapter_repo)

prompt = "Mara 8 kwa 7 ni ngapi?"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=64)
print(tokenizer.decode(out[0], skip_special_tokens=True))

Version

• v0.01 – initial release