DOCSEARCHbyAlgolia
Submit Docs

gemma-skills

repository·main·Indexed 20 days ago

https://github.com/google-gemma/gemma-skills

A repository providing specialized capabilities for interacting with Gemma models. It includes gemma-dev for application development and general knowledge, and gemma-trainer for training, fine-tuning, and adapting models using methods such as SFT, DPO, RLHF, and Reward Modeling on local hardware.

Tokens
3.3K
Snippets
7
Records
21
Agent score
72%

What's inside gemma-skills

  2. Choose the right Gemma training method (SFT, DPO, or RM)

    main

    Select a training workflow based on your objective:

    MethodGoalPrerequisitesOutput
    Supervised Fine-Tuning (SFT)Teaching new domains, specialized tasks, or custom structures.Raw text, instruction pairs, or chat logs.Adaptor trained on prompt/completion pairs.
    Direct Preference Optimization (DPO)Aligning style, behavior, tone, or safety with human preferences.A previously SFT-trained Gemma model and pairwise preference datasets.Aligned model weights (no separate reward head).
    Reward Modeling (RM)Training a scoring system to evaluate response quality.Binary preference pairwise datasets.A classification-style reward head on top of Gemma.

  3. Speed up inference with Multi-Token Prediction (MTP)

    main

    For lower latency, use Multi-Token Prediction (MTP). This uses a lightweight assistant model to propose multiple candidate tokens, which the target model verifies in a single pass.

    Each Gemma 4 target model has a corresponding assistant model following the naming convention <target-model-id>-assistant.

    Assistant Model Repos:

    • google/gemma-4-E2B-it-assistant
    • google/gemma-4-E4B-it-assistant
    • google/gemma-4-12B-it-assistant
    • google/gemma-4-31B-it-assistant
    • google/gemma-4-26B-A4B-it-assistant

  4. Prepare and validate Gemma datasets

    main

    Formatting errors are a primary cause of training failure. Always validate your dataset files using the utility script assets/dataset_prep.py.

    Gemma Chat Prompt Format

    Ensure datasets match the official chat template:

    system
Your instruction here

user
Your query here

model
Your response here

    Tip: Use the tokenizer's apply_chat_template during tokenization to avoid formatting drift.

  5. Perform Direct Preference Optimization (DPO) alignment

    main

    Execute alignment using the assets/dpo_train.py template.

    Critical Rules for DPO:

    1. SFT First: You must perform SFT on the base model using your specific instruction format before running DPO. Running DPO directly on an out-of-domain base model typically fails or degrades formatting.
    2. Beta Parameter: Set beta (DPO temperature) to 0.1. Values between 0.1 and 0.5 control how strictly the model adheres to the reference policy.

  6. Install gemma-skills using Vercel skills CLI

    main

    You can manage and install skills using the Vercel skills CLI via npx.

    To interactively browse and select skills from this repository, use the --list flag. To install a specific skill globally, use the --skill flag followed by the skill name.

    # Interactively browse and install skills.
npx skills add google-gemma/gemma-skills --list

# Install a specific skill (e.g., gemma-dev).
npx skills add google-gemma/gemma-skills --skill gemma-dev --global

  7. Fine-tune Gemma 4 for Vision (Multimodal SFT)

    main

    Gemma 4 models (E2B/E4B/12B/26B/31B) support multimodal vision fine-tuning. Use standard Hugging Face SFTTrainer with a custom visual data collator.

    Dataset Format: Prepare a dataset containing local image paths or PIL images within the message structure:

    {
  "messages": [
    {"role": "user", "content": [
        {"type": "image", "url": "path/to/image.png"},
        {"type": "text", "text": "Describe this image."}
    ]},
    {"role": "assistant", "content": [
        {"type": "text", "text": "An abstract oil painting with vibrant warm gradients."}
    ]}
  ]
}

  8. Fine-tune Gemma 4 for Audio (Multimodal SFT)

    main

    Gemma 4 models (E2B/E4B/12B) support multimodal audio fine-tuning. Feed raw audio arrays (sampled at 16kHz) through the model processor to produce input_features.

    Dataset Format: Maintain conversational formatting, replacing the image type with audio type:

    {
  "messages": [
    {"role": "user", "content": [
        {"type": "text", "text": "Describe this audio."},
        {"type": "audio", "url": "path/to/audio.wav"}
    ]},
    {"role": "assistant", "content": [
        {"type": "text", "text": "This is an audio file of a bird chirping."}
    ]}
  ]
}

  11. Convert trained models to GGUF format

    main

    After LoRA training, convert your model to GGUF for inference.

    Option 1: Native Export via Unsloth (Recommended)

    If you used Unsloth for training, you can export directly to GGUF natively. This handles both merging and quantization automatically.

    Option 2: Manual Conversion with llama.cpp

    If you did not use Unsloth, manually convert your merged Hugging Face model directory using llama.cpp.

  12. Perform Supervised Fine-Tuning (SFT) with QLoRA

    main

    Launch a local QLoRA fine-tuning session using the assets/sft_train.py script.

    Recommended LoRA Hyperparameters:

    • Rank (r): 16 or 32.
    • Alpha (lora_alpha): 32 or 64 (Target lora_alpha = 2 * r).
    • Dropout (lora_dropout): 0.05 or 0.1.
    • Target Modules: Use PEFT's Gemma 4 defaults (targets LM layers).
    • Learning Rate: 2e-4 for QLoRA; 2e-5 for full fine-tuning.