初始化项目，由ModelHub XC社区提供模型

Model: ayh015/myLightningOPD
Source: Original Platform

configs/sft/data/sft_data/openthoughts3_300k_qwen3-8b.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:66cf6cffd11dd523d28cd449afcd5500c62ab47e6617a977039e931d22ae2689
+size 13944494336

configs/sft/dataset_info.json

@@ -0,0 +1,23 @@
+{
+  "openthoughts3_300k_qwen3-8b": {
+    "file_name": "data/sft_data/openthoughts3_300k_qwen3-8b.jsonl",
+    "formatting": "sharegpt",
+    "columns": {
+      "messages": "messages"
+    },
+    "tags": {
+      "role_tag": "role",
+      "content_tag": "content",
+      "user_tag": "user",
+      "assistant_tag": "assistant",
+      "system_tag": "system"
+    }
+  },
+  "openthoughts3_300k_qwen3-32b": {
+    "file_name": "data/sft_data/openthoughts3_300k_qwen3-32b.parquet",
+    "formatting": "sharegpt",
+    "columns": {
+      "messages": "messages"
+    }
+  }
+}

configs/sft/dataset_info_ori_Fri_5_Jun.json

@@ -0,0 +1,16 @@
+{
+  "openthoughts3_300k_qwen3-8b": {
+    "file_name": "data/sft_data/openthoughts3_300k_qwen3-8b.jsonl",
+    "formatting": "sharegpt",
+    "columns": {
+      "messages": "messages"
+    }
+  },
+  "openthoughts3_300k_qwen3-32b": {
+    "file_name": "data/sft_data/openthoughts3_300k_qwen3-32b.parquet",
+    "formatting": "sharegpt",
+    "columns": {
+      "messages": "messages"
+    }
+  }
+}

configs/sft/qwen3-4b-base-open-thoughts3-qwen3-8b.yaml

@@ -0,0 +1,42 @@
+### model
+model_name_or_path: model_weights/qwen3-4b-base
+
+### method
+stage: sft
+do_train: true
+finetuning_type: full
+deepspeed: examples/deepspeed/ds_z0_config.json 
+enable_liger_kernel: true
+packing: true
+
+
+### dataset
+dataset: openthoughts3_300k_qwen3-8b
+template: qwen3
+cutoff_len: 16384
+overwrite_cache: true
+preprocessing_num_workers: 16
+dataloader_persistent_workers: true
+dataloader_pin_memory: true
+dataloader_num_workers: 4
+
+### output
+logging_steps: 1
+save_steps: 100
+save_total_limit: 10
+plot_loss: true
+overwrite_output_dir: false
+save_only_model: false
+#report_to: wandb
+report_to: none
+run_name: qwen3-4b-base-open-thoughts3-qwen3-8b
+
+### train
+per_device_train_batch_size: 4
+gradient_accumulation_steps: 2
+learning_rate: 0.00008
+max_steps: 3000
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000

configs/sft/qwen3-8b-base-open-thoughts3-qwen3-32b.yaml

@@ -0,0 +1,41 @@
+### model
+model_name_or_path: Qwen/Qwen3-8B-Base
+
+### method
+stage: sft
+do_train: true
+finetuning_type: full
+deepspeed: examples/deepspeed/ds_z1_config.json 
+enable_liger_kernel: true
+packing: true
+
+
+### dataset
+dataset: openthoughts3_300k_qwen3-32b
+template: qwen3
+cutoff_len: 16384
+overwrite_cache: true
+preprocessing_num_workers: 16
+dataloader_persistent_workers: true
+dataloader_pin_memory: true
+dataloader_num_workers: 4
+
+### output
+logging_steps: 1
+save_steps: 100
+save_total_limit: 10
+plot_loss: true
+overwrite_output_dir: false
+save_only_model: false
+report_to: wandb
+run_name: qwen3-8b-base-open-thoughts3-qwen3-32b
+
+### train
+per_device_train_batch_size: 2
+gradient_accumulation_steps: 2
+learning_rate: 8e-5
+max_steps: 3000
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000

configs/sft/run_sft.sh

@@ -0,0 +1,49 @@
+#!/usr/bin/env bash
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+# Step 2: Run SFT training with LlamaFactory.
+#
+# Required environment variables:
+#   CONFIG_YAML - Name of the SFT config file in configs/sft/ (e.g. qwen3-4b-base-sft-qwen3-8b.yaml)
+#   OUTPUT_DIR  - Directory for the SFT checkpoint output
+#
+# Optional:
+#   NUM_NODES   - Number of nodes (default: 4)
+#   NUM_GPUS    - GPUs per node (default: 8)
+#   MASTER_ADDR - Master node address (default: localhost)
+#
+# Prerequisites:
+#   - LlamaFactory installed (pip install llamafactory)
+#   - SFT data generated by Step 1 and registered in LlamaFactory's dataset_info.json
+
+set -euo pipefail
+
+: "${CONFIG_YAML:?Set CONFIG_YAML (e.g. qwen3-4b-base-sft-qwen3-8b.yaml)}"
+: "${OUTPUT_DIR:?Set OUTPUT_DIR for SFT checkpoint output}"
+
+NUM_NODES="${NUM_NODES:-4}"
+NUM_GPUS="${NUM_GPUS:-8}"
+MASTER_ADDR="${MASTER_ADDR:-localhost}"
+MASTER_PORT="${MASTER_PORT:-29500}"
+
+# torchrun \
+#     --nnodes "${NUM_NODES}" \
+#     --nproc_per_node="${NUM_GPUS}" \
+#     --rdzv_id $RANDOM \
+#     --rdzv_backend c10d \
+#     --rdzv_endpoint "${MASTER_ADDR}:29500" \
+#     -m llamafactory.cli.train \
+#     "configs/sft/${CONFIG_YAML}" \
+#     "dataset_dir=configs/sft" \
+#     "output_dir=${OUTPUT_DIR}"
+
+FORCE_TORCHRUN=1 \
+NNODES="${NUM_NODES}" \
+NPROC_PER_NODE="${NUM_GPUS}" \
+MASTER_ADDR="${MASTER_ADDR}" \
+MASTER_PORT="${MASTER_PORT}" \
+llamafactory-cli train \
+    "configs/sft/${CONFIG_YAML}" \
+    "dataset_dir=configs/sft" \
+    "output_dir=${OUTPUT_DIR}"