Add Python-code LoRA adapter for gemma-4-e4b-4bit

README.md

@@ -0,0 +1,109 @@
+---
+license: apache-2.0
+license_link: https://ai.google.dev/gemma/docs/gemma_4_license
+base_model: mlx-community/gemma-4-e4b-4bit
+library_name: mlx
+pipeline_tag: text-generation
+inference: false
+tags:
+  - mlx
+  - lora
+  - gemma
+  - code-generation
+  - text-generation
+datasets:
+  - iamtarun/python_code_instructions_18k_alpaca
+---
+
+# gemma-4-e4b-python-lora
+
+A LoRA adapter that fine-tunes [`mlx-community/gemma-4-e4b-4bit`](https://huggingface.co/mlx-community/gemma-4-e4b-4bit)
+for Python code generation, trained with [`mlx_lm`](https://github.com/ml-explore/mlx-lm).
+
+This repo contains **only the adapter weights** — you load it on top of the base
+model, you don't need to download a separate fused model.
+
+## Training details
+
+| | |
+|---|---|
+| Base model | `mlx-community/gemma-4-e4b-4bit` |
+| Dataset | [`iamtarun/python_code_instructions_18k_alpaca`](https://huggingface.co/datasets/iamtarun/python_code_instructions_18k_alpaca) |
+| Fine-tune type | LoRA |
+| LoRA rank | 8 |
+| LoRA scale | 20 |
+| LoRA dropout | 0.0 |
+| Layers adapted | 16 |
+| Learning rate | 1e-5 |
+| Optimizer | Adam |
+| Iterations | 500 |
+
+Training examples were formatted as:
+
+```
+### User:
+{instruction}
+
+### Assistant:
+{output}
+```
+
+## Usage
+
+`mlx_lm.load` needs the adapter on disk, so download it locally first, then load it
+exactly as you would a local adapter:
+
+```bash
+pip install -U mlx-lm huggingface_hub
+```
+
+```python
+from huggingface_hub import snapshot_download
+from mlx_lm import load, generate
+
+adapter_path = snapshot_download("Prashanth-24/gemma-4-e4b-python-lora")
+model, tokenizer = load("mlx-community/gemma-4-e4b-4bit", adapter_path=adapter_path)
+
+prompt = "### User:\nwrite a python code to ADD two numbers\n\n### Assistant:\n"
+print(generate(model, tokenizer, prompt=prompt, max_tokens=200))
+```
+
+## Example output
+
+Prompt: `write a python code to ADD two numbers`
+
+**Base model**
+
+```
+Here is your code:
+
+### response:
+The code to add two numbers is:
+
+### Response:
+Here is your response:
+The response to your query is:
+...
+```
+
+**Fine-tuned model**
+
+```python
+def add_numbers(a, b):
+    if b == 0:
+        return a
+    else:
+        return a / b
+```
+
+The base model loops into repetitive, non-answering text. The fine-tuned model
+reliably switches into clean, structured Python — the LoRA clearly taught the
+`### Assistant:` turn to produce code instead of chatter.
+
+## Limitations
+
+This is a small run (500 iterations, rank 8). Formatting and "answer with code"
+behavior improved a lot, but generalization is imperfect — in the example above the
+fine-tuned model actually emits a *division* function with a zero-check, not a sum,
+for an "add two numbers" prompt. Treat outputs as a draft to review, not
+production-ready code.

adapter_config.json

@@ -0,0 +1,41 @@
+{
+    "adapter_path": "./adapters_python",
+    "batch_size": 1,
+    "clear_cache_threshold": 0,
+    "config": null,
+    "data": "./data_python_alpaca",
+    "fine_tune_type": "lora",
+    "grad_accumulation_steps": 1,
+    "grad_checkpoint": false,
+    "iters": 500,
+    "learning_rate": 1e-05,
+    "lora_parameters": {
+        "rank": 8,
+        "dropout": 0.0,
+        "scale": 20.0
+    },
+    "lr_schedule": null,
+    "mask_prompt": false,
+    "max_seq_length": 1024,
+    "model": "mlx-community/gemma-4-e4b-4bit",
+    "num_layers": 16,
+    "optimizer": "adam",
+    "optimizer_config": {
+        "adam": {},
+        "adamw": {},
+        "muon": {},
+        "sgd": {},
+        "adafactor": {}
+    },
+    "project_name": null,
+    "report_to": null,
+    "resume_adapter_file": null,
+    "save_every": 100,
+    "seed": 0,
+    "steps_per_eval": 50,
+    "steps_per_report": 10,
+    "test": false,
+    "test_batches": 500,
+    "train": true,
+    "val_batches": 25
+}

adapters.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9579543548bf8a5110569233b1a08ed8a597cfbd06b7ce33a690a7ae45096166
+size 30935924