commit 0e58d36bcb7967ab3d6eefdc29fbbaabb33b7cbe
Author: Alan <alan.d@gfp.com.au>
Date:   Wed Mar 25 18:24:39 2026 +1100

    init

diff --git a/.envrc b/.envrc
new file mode 100644
index 0000000..1d953f4
--- /dev/null
+++ b/.envrc
@@ -0,0 +1 @@
+use nix
diff --git a/.gitignore b/.gitignore
new file mode 100644
index 0000000..90eafaf
--- /dev/null
+++ b/.gitignore
@@ -0,0 +1,2 @@
+unsloth_compiled_cache
+outputs_qwen35
diff --git a/shell.nix b/shell.nix
new file mode 100644
index 0000000..c95210a
--- /dev/null
+++ b/shell.nix
@@ -0,0 +1,21 @@
+{nixpkgs ? import <nixpkgs> {}}:
+nixpkgs.mkShell {
+  nativeBuildInputs = with nixpkgs; [
+    ruff
+    python3
+  ];
+
+  LD_LIBRARY_PATH = "${nixpkgs.stdenv.cc.cc.lib}/lib:${nixpkgs.zstd.out}/lib:${nixpkgs.zlib.out}/lib";
+  shellHook = ''
+    if [[ ! -d ".venv" ]]; then
+      python -m venv .venv
+      source .venv/bin/activate
+      pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm7.2/ --upgrade --force-reinstall
+      pip install "unsloth[amd] @ git+https://github.com/unslothai/unsloth"
+      pip install "unsloth-zoo[main] @ git+https://github.com/unslothai/unsloth-zoo"
+      pip install marimo ipython
+    else
+      source .venv/bin/activate
+    fi
+  '';
+}
diff --git a/train.py b/train.py
new file mode 100644
index 0000000..8b1aad1
--- /dev/null
+++ b/train.py
@@ -0,0 +1,91 @@
+import marimo
+
+__generated_with = "0.21.1"
+app = marimo.App(width="full")
+
+
+@app.cell
+def _():
+    from unsloth import FastLanguageModel
+    import torch
+    from datasets import load_dataset
+    from trl import SFTTrainer, SFTConfig
+
+    return FastLanguageModel, SFTConfig, SFTTrainer, load_dataset
+
+
+@app.cell
+def _(load_dataset):
+    max_seq_length = 2048  # start small; scale up after it works
+
+    # Example dataset (replace with yours). Needs a "text" column.
+    url = "https://huggingface.co/datasets/laion/OIG/resolve/main/unified_chip2.jsonl"
+    dataset = load_dataset("json", data_files={"train": url}, split="train")
+    return dataset, max_seq_length
+
+
+@app.cell
+def _(dataset):
+    dataset.to_pandas().head(50)
+    return
+
+
+@app.cell
+def _(FastLanguageModel, max_seq_length):
+    # unsloth/Qwen3.5-0.8B
+    model, tokenizer = FastLanguageModel.from_pretrained(
+        model_name = "unsloth/Qwen3.5-0.8B",
+        max_seq_length = max_seq_length,
+        load_in_4bit = False,     # MoE QLoRA not recommended, dense 27B is fine
+        load_in_16bit = True,     # bf16/16-bit LoRA
+        full_finetuning = False,
+    )
+
+    model = FastLanguageModel.get_peft_model(
+        model,
+        r = 16,
+        target_modules = [
+            "q_proj", "k_proj", "v_proj", "o_proj",
+            "gate_proj", "up_proj", "down_proj",
+        ],
+        lora_alpha = 16,
+        lora_dropout = 0,
+        bias = "none",
+        # "unsloth" checkpointing is intended for very long context + lower VRAM
+        use_gradient_checkpointing = "unsloth",
+        random_state = 3407,
+        max_seq_length = max_seq_length,
+    )
+    return model, tokenizer
+
+
+@app.cell
+def _(SFTConfig, SFTTrainer, dataset, max_seq_length, model, tokenizer):
+    trainer = SFTTrainer(
+        model = model,
+        train_dataset = dataset,
+        tokenizer = tokenizer,
+        args = SFTConfig(
+            max_seq_length = max_seq_length,
+            per_device_train_batch_size = 1,
+            gradient_accumulation_steps = 4,
+            warmup_steps = 10,
+            max_steps = 100,
+            logging_steps = 1,
+            output_dir = "outputs_qwen35",
+            optim = "adamw_8bit",
+            seed = 3407,
+            dataset_num_proc = 1,
+        ),
+    )
+    return (trainer,)
+
+
+@app.cell
+def _(trainer):
+    trainer.train()
+    return
+
+
+if __name__ == "__main__":
+    app.run()