charbull/ppo-gemma-lora-only-v1

---
license: mit
base_model: mlx-community/gemma-3-1b-it-bf16
---

# LoRA Adapter for `mlx-community/gemma-3-1b-it-bf16`

This repository contains a LoRA adapter trained to play Wordle.

## Usage

This adapter should be used with the base model `mlx-community/gemma-3-1b-it-bf16`.
You will need to define the `LoRALinear` class provided in the original MLX examples.

```python
import mlx.core as mx
from mlx_lm import load
from huggingface_hub import hf_hub_download

# You must have the LoRALinear class defined in your script.
# (Insert the LoRALinear class definition here)

# Load the base model
model, tokenizer = load("mlx-community/gemma-3-1b-it-bf16")

# Apply the LoRA layers to the model
# This must be done *before* loading the adapter weights
model.freeze()
for l in model.model.layers[-16:]:
    l.self_attn.q_proj = LoRALinear.from_linear(l.self_attn.q_proj, rank=8)
    l.self_attn.v_proj = LoRALinear.from_linear(l.self_attn.v_proj, rank=8)
    if hasattr(l, "block_sparse_moe"):
        l.block_sparse_moe.gate = LoRALinear.from_linear(l.block_sparse_moe.gate, rank=8)

# Download and load the adapter weights
adapter_path = hf_hub_download(repo_id="charbull/ppo-gemma-lora-only-v1", filename="wordle_gemma3_grpo_adapters.npz")
weights = mx.load(adapter_path)
model.update(weights) # Note: mlx-lm > 0.1.1 has model.update_from_safetensors

print("Adapter loaded successfully.")
# The model is now ready for generation or fusion.