first commit

.mv

@@ -0,0 +1 @@
+Revision:master,CreatedAt:1732026261

README.md

@@ -1,3 +1,95 @@
-# Steel-LLM_a13737944229998592552118
+---
+license: apache-2.0
+---
+<div align="center">
 
-Steel-LLM
+# 开源中文预训练语言模型Steel-LLM
+由zhanshijin和lishu14创建
+</div>
+
+
+## 👋 介绍
+Steel-LLM是个人发起的利用业余时间从零开始预训练中文大模型的项目。我们使用了1T+的数据预训练一个1B左右参数量的中文LLM，耗时8个月。我们分享了数据收集、数据处理、预训练框架修改、模型设计、模型微调等全过程，并开源全部代码。让每个人在有8~几十张卡的情况下都能复现我们的工作。得益于开源中文数据，Steel LLM在中文benchmark上表现优于一些大几倍的机构发布的LLM，最终在ceval达到了38分，cmmlu达到了33分。
+<p align="center">
+        🐱 <a href="https://github.com/zhanshijinwat/Steel-LLM">Github</a>&nbsp&nbsp 
+        &nbsp&nbsp 📑 <a href="https://www.zhihu.com/people/zhan-shi-jin-27">Blog</a> &nbsp&nbsp🌐公众号：炼钢AI
+
+"Steel(钢)"取名灵感来源于华北平原一只优秀的乐队“万能青年旅店（万青）”。乐队在做一专的时候条件有限，自称是在“土法炼钢”，但却是一张神专。我们训练LLM的条件同样有限，但也希望能炼出好“钢”来。
+
+## 📖 预训练数据
+预训练数据方面，Steel-LLM主要使用了wanjuan1.0、Skywork/Skypile-150B数据集、starcoder的python/java/c++数据。另外也加入了中文维基百科、百度百科、知乎问答等数据，转换为token id后占用1.7T硬盘空间。Steel-LLM也对问答数据以及代码数据使用data-juicer进行了数据清洗，数据收集及数据处理的具体细节见我的博客：
+https://mp.weixin.qq.com/s/yqmtHLuuNV9075qHgzhcPw
+
+## 🎰 训练框架
+训练框架方面，我们修改了TinyLlama训练程序，兼容Hugginface格式模型、支持了数据断点续训、支持了追加新的数据等能力。训练前20k checkpoint使用8 * A100，之后使用的是8 * H800
+具体的技术细节见我的博客：https://mp.weixin.qq.com/s/KPRir6bK3MZZ-vMFTfhUQQ
+
+## 🤖模型结构
+
+tokenizer方面，使用了Qwen/Qwen1.5-MoE-A2.7B-Chat的tokenizer。模型结构方面基于Qwen1.5模型进行了以下的新结构的尝试：
+- FFN层使用softmax moe，相同参数量下有更高的训练速度
+- 使用双层的SwiGLU
+具体的技术细节见我的博客：https://mp.weixin.qq.com/s/JaZyf1jOEOtNDCcFqSj8TQ
+
+## 💡 微调
+微调阶段主要使用了BAAI/Infinity-Instruct、预训练数据中的wanjuan中文选择题部分（回炉重造）、ruozhiba等数据。尝试了COT/非COT微调、刷榜测试。
+具体的实验细节见我的博客：https://mp.weixin.qq.com/s/KK0G0spNw0D9rPUESkHMew
+
+## 🏅 评估
+Steel-LLM在CEVAL和CMMLU上进行了测试。Steel-LLM旨在训练一个中文LLM，80%的训练数据都是中文，因此并没有在英文benchmark上进行评测。
+其他模型的指标来自于CEVAL论文、MiniCPM技术报告、MAP-Neo技术报告等途径。更多模型的指标可查看之前的<a href=https://mp.weixin.qq.com/s/KK0G0spNw0D9rPUESkHMew>博客</a>
+
+|                              | CEVAL  | CMMLU |
+|------------------------------|--------|-------|
+| Steel-LLM                    | 38.57  | 33.48 |
+| Tiny-Llama-1.1B              | 25.02  | 24.03 |
+| Gemma-2b-it                  | 32.3   | 33.07 |
+| Phi2(2B)	                    | 23.37	 | 24.18 |
+| Deepseek-coder-1.3B-instruct |  28.33 | 27.75 |
+| CT-LLM-SFT-2B                | 41.54  | 41.48 |
+| MiniCPM-2B-sft-fp32          | 49.14  | 51.0  |
+| Qwen1.5-1.8B-Chat            | 56.84  | 54.11 |
+| ChatGLM-6B                   | 38.9   | -     |
+| Moss                         | 33.1   | -     |
+| LLAMA-65B                    | 34.7   | -     |
+| Qwen-7B                      | 58.96  | 60.35 |
+| Gemma-7B                     | 42.57  | 44.20 |
+| OLMo-7B                      | 35.18  | 35.55 |
+| MAP-NEO-7B                   | 56.97  | 55.01 |
+
+## ⛏️ 快速使用
+```python
+from modelscope import AutoModelForCausalLM, AutoTokenizer
+
+model_name = "zhanshijin/Steel-LLM"
+
+model = AutoModelForCausalLM.from_pretrained(
+    model_name,
+    torch_dtype="auto",
+    device_map="auto"
+)
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+
+prompt = "你是谁开发的"
+messages = [
+    {"role": "user", "content": prompt}
+]
+text = tokenizer.apply_chat_template(
+    messages,
+    tokenize=False,
+    add_generation_prompt=True
+)
+model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
+
+generated_ids = model.generate(
+    **model_inputs,
+    max_new_tokens=512
+)
+generated_ids = [
+    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
+]
+
+response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+print(response)
+
+```

added_tokens.json

@@ -0,0 +1,5 @@
+{
+  "<|endoftext|>": 151643,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644
+}

config.json

@@ -0,0 +1,38 @@
+{
+  "FFN_type": "softmoe_v3",
+  "_name_or_path": "/data/model/llm/hf_model/steel-llm-step-1060000-ckpt",
+  "architectures": [
+    "SteelForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_steel.SteelConfig",
+    "AutoModel": "modeling_steel.SteelForCausalLM",
+    "AutoModelForCausalLM": "modeling_steel.SteelForCausalLM"
+  },
+  "bos_token_id": 151643,
+  "eos_token_id": 151645,
+  "hidden_act": "silu",
+  "hidden_size": 1792,
+  "initializer_range": 0.02,
+  "intermediate_size": 1792,
+  "max_position_embeddings": 32768,
+  "max_window_layers": 21,
+  "mlp_div_ratio": 4,
+  "mlp_type": "senet",
+  "model_type": "qwen2",
+  "n_experts": 6,
+  "num_attention_heads": 32,
+  "num_hidden_layers": 18,
+  "num_key_value_heads": 32,
+  "rms_norm_eps": 1e-06,
+  "rope_theta": 1000000.0,
+  "sliding_window": 32768,
+  "slots_per_expert": 1,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float32",
+  "transformers_version": "4.43.4",
+  "use_cache": false,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

configuration_steel.py

@@ -0,0 +1,131 @@
+# coding=utf-8
+# Copyright 2024 zhanshijin and lishu. All rights reserved.
+#
+# This project copy from qwen1.5
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Steel model configuration"""
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+
+
+class SteelConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Qwen2Model`]. It is used to instantiate a
+    Qwen2 model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of
+    Qwen2-7B-beta [Qwen/Qwen2-7B-beta](https://huggingface.co/Qwen/Qwen2-7B-beta).
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 151936):
+            Vocabulary size of the Qwen2 model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`Qwen2Model`]
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 22016):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_key_value_heads (`int`, *optional*, defaults to 32):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `32`.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 32768):
+            The maximum sequence length that this model might ever be used with.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied.
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        use_sliding_window (`bool`, *optional*, defaults to `False`):
+            Whether to use sliding window attention.
+        sliding_window (`int`, *optional*, defaults to 4096):
+            Sliding window attention (SWA) window size. If not specified, will default to `4096`.
+        max_window_layers (`int`, *optional*, defaults to 28):
+            The number of layers that use SWA (Sliding Window Attention). The bottom layers use SWA while the top use full attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+    """
+
+    model_type = "qwen2"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=151936,
+        hidden_size=4096,
+        intermediate_size=22016,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=32,
+        hidden_act="silu",
+        max_position_embeddings=32768,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        use_sliding_window=False,
+        sliding_window=4096,
+        max_window_layers=28,
+        attention_dropout=0.0,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.use_sliding_window = use_sliding_window
+        self.sliding_window = sliding_window
+        self.max_window_layers = max_window_layers
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.attention_dropout = attention_dropout
+
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 151643,
+  "eos_token_id": 151645,
+  "transformers_version": "4.43.4"
+}

softmoe_v3.py

@@ -0,0 +1,112 @@
+from typing import Callable
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+def softmax(x: torch.Tensor, dim: int | tuple[int, ...]) -> torch.Tensor:
+    """
+    Compute the softmax along the specified dimensions.
+    This function adds the option to specify multiple dimensions
+
+    Args:
+        x (torch.Tensor): Input tensor.
+        dims (int or tuple[int]): The dimension or list of dimensions along which the softmax probabilities are computed.
+
+    Returns:
+        torch.Tensor: Output tensor containing softmax probabilities along the specified dimensions.
+    """
+    dtype = x.dtype
+    x = x.to(torch.float32)
+    max_vals = torch.amax(x, dim=dim, keepdim=True)
+    e_x = torch.exp(x - max_vals)
+    sum_exp = e_x.sum(dim=dim, keepdim=True)
+    return (e_x / sum_exp).to(dtype)
+
+# copy from https://github.com/bwconrad/soft-moe
+class SteelSoftMoEV3(nn.Module):
+    """
+    A wrapper class to create a Soft Mixture of Experts layer.
+
+    From "From Sparse to Soft Mixtures of Experts"
+    https://arxiv.org/pdf/2308.00951.pdf
+    """
+
+    def __init__(
+        self,
+        config,
+        layer: Callable,
+    ) -> None:
+        """
+        Args:
+            dim (int): Dimensionality of input features.
+            num_experts (int): Number of experts.
+            slots_per_expert (int): Number of token slots per expert.
+            layer (Callable): Network layer of the experts.
+            normalize (bool): Normalize input and phi (sec. 2.3 from paper)
+            **layer_kwargs: Additional keyword arguments for the layer class.
+        """
+        super().__init__()
+
+        self.dim = config.hidden_size
+        self.num_experts = config.n_experts
+        self.slots_per_expert = config.slots_per_expert if hasattr(config, "slots_per_expert") else 1
+        self.normalize = True
+
+        # Initialize phi and normalization scaling factor
+        self.phi = nn.Parameter(torch.zeros(self.dim, self.num_experts, self.slots_per_expert))
+        if self.normalize:
+            self.scale = nn.Parameter(torch.ones(1))
+
+        # Initialize phi using LeCun normal initialization
+        # https://github.com/google-research/vmoe/blob/662341d007650d5bbb7c6a2bef7f3c759a20cc7e/vmoe/projects/soft_moe/router.py#L49C1-L49C1
+        nn.init.normal_(self.phi, mean=0, std=1 / self.dim**0.5)
+
+        # Create a list of expert networks
+        self.experts = nn.ModuleList(
+            [layer(config) for _ in range(self.num_experts)]
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Forward pass through the Soft-MoE layer (algorithm 1 from paper).
+
+        Args:
+            x (torch.Tensor): Input tensor of shape [batch_size, seq_len, input_dim].
+
+        Returns:
+            torch.Tensor: Output tensor of shape [batch_size, seq_len, input_dim].
+        """
+        assert (
+            x.shape[-1] == self.dim
+        ), f"Input feature dim of {x.shape[-1]} does not match layer dim of {self.dim}"
+        assert (
+            len(x.shape) == 3
+        ), f"Input expected to have 3 dimensions but has {len(x.shape)}"
+
+        phi = self.phi
+
+        # Normalize input and phi
+        if self.normalize:
+            x = F.normalize(x, dim=2)  # [b, m, d]
+            phi = self.scale * F.normalize(phi, dim=0)  # [d, n, p]
+
+        # Compute dispatch and combine weights
+        logits = torch.einsum("bmd,dnp->bmnp", x, phi)
+        d = softmax(logits, dim=1)
+        c = softmax(logits, dim=(2, 3))
+        # tmp = c[0,:,:,0].reshape([c.shape[1],-1])
+        # print("num:",tmp, "shape:",tmp.shape, "sum:",tmp.sum(dim=1))
+        # Compute input slots as weighted average of input tokens using dispatch weights
+        xs = torch.einsum("bmd,bmnp->bnpd", x, d)
+
+        # Apply expert to corresponding slots
+        ys = torch.stack(
+            [f_i(xs[:, i, :, :]) for i, f_i in enumerate(self.experts)], dim=1
+        )
+
+        # Compute output tokens as weighted average of output slots using combine weights
+        y = torch.einsum("bnpd,bmnp->bmd", ys, c)
+
+        return y

special_tokens_map.json

@@ -0,0 +1,20 @@
+{
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>"
+  ],
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer_config.json

@@ -0,0 +1,44 @@
+{
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>"
+  ],
+  "bos_token": null,
+  "chat_template": "{% set system_message = 'You are a helpful assistant.' %}{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% else %}{% set loop_messages = messages %}{% endif %}{% if system_message is defined %}{{ '<|im_start|>system\n' + system_message + '<|im_end|>\n' }}{% endif %}{% for message in loop_messages %}{% set content = message['content'] %}{% if message['role'] == 'user' %}{{ '<|im_start|>user\n' + content + '<|im_end|>\n<|im_start|>assistant\n' }}{% elif message['role'] == 'assistant' %}{{ content + '<|im_end|>' + '\n' }}{% endif %}{% endfor %}",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "model_max_length": 32768,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}