task-decomposer-270m

Gemma3 270M - Task Decomposer

A fine-tuned version of Gemma3 270M optimized for decomposing tasks into actionable subtasks with time estimates.

Model Details

• Base Model: unsloth/gemma-3-270m-it
• Fine-tuning Method: LoRA (r=128, alpha=128) with Unsloth
• Training Data: 804 diverse task decomposition examples
• Format: GGUF Q8_0 quantization (~270MB)
• Training Duration: 3 epochs (~20-30 minutes on T4 GPU)
• Use Case: Breaking down everyday tasks into structured subtasks
• Optimized For: Mobile deployment via llama.cpp / llama_sdk

Training Data Categories

The model was trained on 804 examples across 11 diverse categories:

Total: 804 examples (723 training, 81 evaluation)

Usage

Input Format

Decompose this task into subtasks: [YOUR TASK HERE]

Expected Output Format

The model outputs JSON with subtasks containing:

• title: Brief, actionable subtask description
• estimateMinutes: Conservative time estimate in minutes
• order: Sequential order (1, 2, 3, ...)

[
  {"title": "Choose party date and time", "estimateMinutes": 15, "order": 1},
  {"title": "Create guest list", "estimateMinutes": 20, "order": 2},
  {"title": "Book venue or prepare home", "estimateMinutes": 45, "order": 3},
  {"title": "Order cake and plan menu", "estimateMinutes": 30, "order": 4},
  {"title": "Send invitations", "estimateMinutes": 20, "order": 5},
  {"title": "Shop for decorations", "estimateMinutes": 60, "order": 6}
]

With llama.cpp

./llama-cli -m taskapp-gemma3-270m-Q8_0.gguf \
    -p "Decompose this task into subtasks: Clean the garage" \
    --temp 0.7 \
    --top-p 0.9

With llama_sdk (Flutter/Dart)

final llama = Llama(LlamaController(
  modelPath: 'path/to/taskapp-gemma3-270m-Q8_0.gguf',
  nCtx: 2048,
  nBatch: 512,
  greedy: false,
));

final messages = [
  UserLlamaMessage('Decompose this task into subtasks: Plan birthday party'),
];

final stream = llama.prompt(messages);
await for (final response in stream) {
  print(response.text);
}

With Python (transformers)

from transformers import AutoTokenizer, AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("johkamb/task-decomposer-270m")
tokenizer = AutoTokenizer.from_pretrained("johkamb/task-decomposer-270m")

prompt = "Decompose this task into subtasks: Organize garage"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0]))

Example Outputs

Example 1: Plan Birthday Party

[
  {"title": "Choose date and send save-the-dates", "estimateMinutes": 15, "order": 1},
  {"title": "Create guest list", "estimateMinutes": 20, "order": 2},
  {"title": "Book venue or prepare home", "estimateMinutes": 45, "order": 3},
  {"title": "Order cake and plan menu", "estimateMinutes": 30, "order": 4},
  {"title": "Send formal invitations", "estimateMinutes": 20, "order": 5},
  {"title": "Shop for decorations and party supplies", "estimateMinutes": 60, "order": 6}
]

Example 2: Write Research Paper

[
  {"title": "Choose topic and create thesis", "estimateMinutes": 45, "order": 1},
  {"title": "Research and gather sources", "estimateMinutes": 120, "order": 2},
  {"title": "Create detailed outline", "estimateMinutes": 30, "order": 3},
  {"title": "Write first draft", "estimateMinutes": 180, "order": 4},
  {"title": "Revise and edit", "estimateMinutes": 90, "order": 5},
  {"title": "Format citations and bibliography", "estimateMinutes": 45, "order": 6}
]

Example 3: Clean Garage

[
  {"title": "Remove everything from garage", "estimateMinutes": 60, "order": 1},
  {"title": "Sweep and clean floor", "estimateMinutes": 20, "order": 2},
  {"title": "Sort items into keep/donate/trash", "estimateMinutes": 45, "order": 3},
  {"title": "Organize items on shelves", "estimateMinutes": 40, "order": 4},
  {"title": "Dispose of unwanted items", "estimateMinutes": 30, "order": 5}
]

Training Details

Hyperparameters

• Base Model: unsloth/gemma-3-270m-it
• Training Examples: 804 (no eval split for full data utilization)
• Epochs: 3
• Batch Size: 4 per device
• Gradient Accumulation: 1 (effective batch size = 4)
• Learning Rate: 5e-5
• LR Scheduler: Linear
• Warmup Steps: 10
• Optimizer: AdamW 8-bit
• Weight Decay: 0.001
• LoRA Rank (r): 128
• LoRA Alpha: 128
• LoRA Dropout: 0
• Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Max Sequence Length: 2048 tokens
• Precision: FP32 (Gemma3 requirement)
• Gradient Checkpointing: unsloth (memory efficient)

Training Environment

• GPU: NVIDIA T4 (Google Colab)
• Training Time: ~20-30 minutes
• Framework: Unsloth + transformers + TRL (SFTConfig)
• Quantization: FP32 during training, Q8_0 for deployment

Performance Metrics

• Final Training Loss: < 0.5 (excellent), < 1.0 (good)
• Trainable Parameters: ~3-5% of total model parameters (r=128)
• Model Size:
  • Original: ~540MB (FP32)
  • Quantized (Q8_0): ~270MB
• Quality: Excellent (Q8_0 has minimal quality loss vs FP32)
• Inference Speed: Fast on mobile devices (optimized by llama.cpp)

Limitations

• Language: Optimized for English only
• Task Scope: Best for tasks completable within days/weeks
• Domain: Optimized for everyday tasks (household, work, school, personal)
• Time Estimates: Conservative estimates (designed to under-promise)
• Specialized Tasks: May struggle with highly technical or niche domains
• JSON Output: Occasionally may need retry for valid JSON formatting
• Context Length: Limited to 2048 tokens (~1500 words)

Ethical Considerations

• Bias: Training data focuses on common Western/American tasks; may not reflect all cultures
• Time Estimates: Conservative estimates may not fit all working styles
• Task Decomposition: Suggested subtasks are general; individual needs may vary
• Not for: Medical advice, legal guidance, or critical decision-making

Acknowledgments

• Base Model: Google Gemma Team
• Fine-tuning: Unsloth AI (optimized training)
• Quantization: llama.cpp team (GGUF format)
• Training Platform: Google Colab