GLM-4.7 training script

train_glm.py

@@ -0,0 +1,86 @@
+# /// script
+# requires-python = ">=3.10"
+# dependencies = [
+#     "trl>=0.12.0",
+#     "peft>=0.7.0",
+#     "transformers>=4.36.0",
+#     "accelerate>=0.24.0",
+#     "trackio",
+#     "datasets",
+# ]
+# ///
+
+"""
+Agent Zero SFT: zai-org/GLM-4.7-Flash (30B MoE)
+LoRA fine-tuning on agent-zero-sft-v1 dataset.
+Router layers frozen - only attention layers trained.
+"""
+
+import trackio
+from datasets import load_dataset
+from peft import LoraConfig
+from trl import SFTTrainer, SFTConfig
+
+# Load dataset
+print("Loading dataset...")
+train_ds = load_dataset("wheattoast11/agent-zero-sft-v1", data_files="data/train.jsonl", split="train")
+val_ds = load_dataset("wheattoast11/agent-zero-sft-v1", data_files="data/validation.jsonl", split="train")
+print(f"Train: {len(train_ds)}, Val: {len(val_ds)}")
+
+config = SFTConfig(
+    output_dir="agent-zero-glm-4.7-v1",
+    push_to_hub=True,
+    hub_model_id="wheattoast11/agent-zero-glm-4.7-v1",
+    hub_strategy="every_save",
+    hub_private_repo=True,
+
+    num_train_epochs=2,
+    per_device_train_batch_size=1,
+    gradient_accumulation_steps=16,
+    learning_rate=1e-4,
+    bf16=True,
+    gradient_checkpointing=True,
+
+    logging_steps=10,
+    save_strategy="steps",
+    save_steps=50,
+    save_total_limit=2,
+
+    eval_strategy="steps",
+    eval_steps=50,
+
+    warmup_ratio=0.1,
+    lr_scheduler_type="cosine",
+
+    report_to="trackio",
+    project="agent-zero-finetune",
+    run_name="glm-4.7-flash-sft-v1",
+)
+
+# LoRA targeting attention layers only (router layers frozen)
+peft_config = LoraConfig(
+    r=16,
+    lora_alpha=32,
+    lora_dropout=0.05,
+    bias="none",
+    task_type="CAUSAL_LM",
+    target_modules=["q_proj", "v_proj", "k_proj", "o_proj"],
+)
+
+print("Initializing trainer...")
+trainer = SFTTrainer(
+    model="zai-org/GLM-4.7-Flash",
+    train_dataset=train_ds,
+    eval_dataset=val_ds,
+    args=config,
+    peft_config=peft_config,
+)
+
+print("Starting training...")
+trainer.train()
+
+print("Pushing to Hub...")
+trainer.push_to_hub()
+
+trackio.finish()
+print("Done! Model at: https://huggingface.co/wheattoast11/agent-zero-glm-4.7-v1")