CausalLM 14B-EXL2
- Model creator: CausalLM
- Original model: CausalLM 14B
Description
Experimental exl2 quantization for CausalLM-14B for Exllamav2.
I had some issues during quantization process, so I suspect it might have quality issues.
3.5bpw version barely fits my 12GB VRAM but has unusually high perplexity for wikitext dataset.
I couldn't measure perplexity for 4bpw version and to compare it with TheBloke's GPTQ, so I have no idea if my quantization has issues or it supposed to be like this.
You could try this exl2 version but I'd recommend to use TheBloke's GPTQ version instead.
How to run
This quantization method uses GPU and requires Exllamav2 loader which can be found in following applications:
Original model card:
Image drawn by GPT-4 DALL·E 3 TL;DR: Perhaps better than all existing models < 70B, in most quantitative evaluations...
CausalLM 14B - Fully Compatible with Meta LLaMA 2
Use the transformers library that does not require remote/external code to load the model, AutoModelForCausalLM and AutoTokenizer (or manually specify LlamaForCausalLM to load LM, GPT2Tokenizer to load Tokenizer), and model quantization is fully compatible with GGUF (llama.cpp), GPTQ, and AWQ.
News: DPO ver. Rank #1 ~13B - SOTA model of its size on 🤗 Open LLM Leaderboard
Recent Updates: DPO-α Version outperforms Zephyr-β in MT-Bench
Friendly reminder: If your VRAM is insufficient, you should use the 7B model instead of the quantized version.
Compared to the quantized versions, the 7B version and the 14B version demonstrate a high level of consistency.
llama.cpp GGUF models GPT2Tokenizer fixed by Kerfuffle on https://github.com/ggerganov/llama.cpp/pull/3743, new models are now reuploaded.
Thanks TheBloke for GGUF quants: https://huggingface.co/TheBloke/CausalLM-14B-GGUF
Caution: Unofficial GPTQ and AWQ models may have issues as they use Wikitext for calibration, while this model has undergone considerable training on a synthesized Wikipedia conversation dataset.
It is not recommended to use any form of quantization, but rather to use smaller-sized models, as the 7B and 14B versions have high consistency. However, if you do use model quantization, please use GGUF.
Read Me:
Also see 7B Version
This model was trained based on the model weights of Qwen (and LLaMA2 was used, yes, for calculating some initial weights), you may also need to comply with the commercial use restrictions of these two models depending on the situation. The training process utilized a model architecture that was identical to LLaMA2, using the same attention calculation method as the original MHA LLaMA2 models, and no additional scaling applied to the Rotary Positional Encoding (RoPE).
We manually curated a SFT dataset of 1.3B tokens for training, utilizing open source datasets from Hugging Face. For most of these sentences, we performed manual or synthetic rewrites and generated alternate language versions using larger language models. Additionally, we conducted augmented text training using carefully selected entries from Wikipedia, as well as featured entries from Fandom and filtered entries from Moegirlpedia. In order to strike a balance between efficiency and quality, 100% of the data used for training was synthetic data, no direct use of text from the internet or original texts from publicly available datasets was employed for fine-tuning.
The 7B version of the model is a distilled version of the 14B model, specifically designed for speculative sampling. Therefore, it is important to exercise caution when directly using the model, as it may produce hallucinations or unreliable outputs.
Please note that the model was trained on unfiltered internet data. Since we do not have the capacity to vet all of it, there may be a substantial amount of objectionable content, pornography, violence, and offensive language present that we are unable to remove. Therefore, you will still need to complete your own checks on the model's safety and filter keywords in the output. Due to computational resource constraints, we are presently unable to implement RLHF for the model's ethics and safety, nor training on SFT samples that refuse to answer certain questions for restrictive fine-tuning.
Bonus: The model underwent some fine-tuning on the prompt format introduced in LLaVA1.5 that is unrelated to image attention calculation. Therefore, aligning the ViT Projection module with frozen LM under visual instructions would enable rapid implementation of effective multimodal capabilities.
PROMPT FORMAT:
System Prompt must not be empty!
MMLU:
stem ACC: 64.19
Humanities ACC: 61.40
other ACC: 71.64
social ACC: 75.37
AVERAGE ACC:67.36 (Outperforms ALL models under 70B, very close to those best 70B fine-tunes)
CEval (Val):
STEM ACC: 66.71
Social Science ACC: 85.10
Humanities ACC: 76.68
Other ACC: 70.23
Hard ACC:54.71
AVERAGE ACC:73.10 (Outperforms Qwen-14B, and GPT-4)
GSM8K
Zero-shot ACC 0.7012888551933283 (Outperforms MetaMath-13B, Qwen-14B)
AlpacaEval Leaderboard
win_rate | standard_error | n_wins | n_wins_base | n_draws | n_total | mode | avg_length | |
---|---|---|---|---|---|---|---|---|
causallm-14b | 88.26087 | 1.116333 | 705 | 89 | 11 | 805 | community | 1391 |
Win rate 88.26% on AlpacaEval Leaderboard view raw
MT-Behch on DPO Version
Model | MT-Bench |
---|---|
GPT-4 | 8.99 |
GPT-3.5-Turbo | 7.94 |
Zephyr-7b-β (Overfitting) | 7.34 |
Zephyr-7b-α | 6.88 |
CausalLM/14B-DPO-α | 7.618868 |
CausalLM/7B-DPO-α | 7.038125 |
Other languages
We are currently unable to produce accurate benchmark templates for non-QA tasks (languages other than English and Chinese). However, we will be working on other language versions of the QA-Task challenge in the near future.
Japanese Benchmark
Task | Version | Metric | Value | Stderr | |
---|---|---|---|---|---|
jcommonsenseqa-1.1-0.6 | 1.1 | acc | 0.8213 | ± | 0.0115 |
JCommonsenseQA benchmark result is very, very close to Japanese Stable LM Gamma 7B (83.47), current SOTA Japanese LM. However, our model was not trained on a particularly large amount of text in Japanese. This seems to reflect the cross-language transferability of metalinguistics.
🤗 Open LLM Leaderboard
SOTA chat model of its size on 🤗 Open LLM Leaderboard.
Dec 3, 2023 DPO Version Rank #1 non-base model, of its size on 🤗 Open LLM Leaderboard, outperforms ALL ~13B chat models.
因果语言模型 14B - 与 Meta LLaMA 2 完全兼容
使用无需远程/外部代码的transformers库加载模型,AutoModelForCausalLM和AutoTokenizer(或者手动指定LlamaForCausalLM加载LM, GPT2Tokenizer加载Tokenizer),并且模型量化与GGUF(llama.cpp)、GPTQ、AWQ完全兼容。
新消息:DPO 版本在~13B排名第1 🤗 Open LLM 排行榜上同尺寸的所有模型中评分最高
最近更新: DPO-α Version 在 MT-Bench 超过 Zephyr-β
友情提示:如果您的显存不足,您应该使用7B模型而不是量化版本。
与量化版本相比,7B 版本和 14B 版本具有高度的一致性。
llama.cpp GGUF models GPT2Tokenizer 支持由 Kerfuffle 修复于 https://github.com/ggerganov/llama.cpp/pull/3743,新模型稍后上传。
感谢 TheBloke 制作 GGUF 版本量化模型: https://huggingface.co/TheBloke/CausalLM-14B-GGUF
注意: 非官方 GPTQ 和 AWQ 模型可能存在问题,因为它们使用 Wikitext 进行校准,而该模型已经在合成的 Wikipedia 对话数据集上经过了大量的训练。
不建议使用任何形式的量化,而是使用较小尺寸的模型,因为7B和14B版本具有较高的一致性。 但是,如果您确实使用模型量化,请使用 GGUF。
请读我:
另请参阅7B版本
该模型是基于Qwen的权重(并使用了LLaMA2权重,是的,用于计算一些权重初始化),您根据情况可能还需要遵守这两个模型的商业使用限制。训练过程中使用了与LLaMA2相同的模型结构,使用原始MHA LLaMA2模型的相同注意力计算方法,对旋转位置编码(RoPE)没有进行额外的缩放。
我们手动筛选了一个包含13亿个标记的SFT数据集进行训练,利用了Hugging Face的开源数据集。对于大多数句子,我们进行了手动或合成改写,并使用更大的语言模型生成了其他语言版本。此外,我们还使用了精心挑选的来自维基百科的条目、来自Fandom的精选条目以及来自萌娘百科的过滤条目进行增强文本训练。为了在效率和质量之间取得平衡,训练所使用的100%数据都是合成数据,没有直接使用来自互联网或公开可用数据集的原始文本进行微调。
7B版本的模型是14B模型的精简版本,专门设计用于推测抽样。因此,在直接使用模型时,需要谨慎行事,因为它可能会产生幻觉或不可靠的输出。
请注意,模型是在未经过滤的互联网数据上进行训练的。由于我们无法审核所有数据,可能会出现大量不良内容、色情、暴力和冒犯性语言,我们无法删除这些内容。因此,您仍然需要对模型的安全性进行自己的检查,并对输出中的关键词进行过滤。由于计算资源的限制,我们目前无法为模型的伦理和安全实施RLHF,也无法对拒绝回答某些问题的SFT样本进行训练以进行限制性微调。
额外奖励:模型在LLaVA1.5中引入的提示格式上进行了一些微调,与图像注意力计算无关。因此,将ViT投影模块与冻结的LM对齐,并根据视觉指令实施快速实现有效的多模态能力。
提示格式:
系统提示不能为空!
MMLU:
STEM准确率:64.19
人文及艺术学科准确率:61.40
其他学科准确率:71.64
社会学科准确率:75.37
平均准确率:67.36(超过所有70B以下的模型,非常接近最佳70B微调模型)
CEval(验证集):
STEM准确率:66.71
社会科学准确率:85.10
人文学科准确率:76.68
其他学科准确率:70.23
困难准确率:54.71
平均准确率:73.10(超过Qwen-14B和GPT-4)
GSM8K
零样本准确率0.7012888551933283(超过MetaMath-13B和Qwen-14B)
AlpacaEval Leaderboard
win_rate | standard_error | n_wins | n_wins_base | n_draws | n_total | mode | avg_length | |
---|---|---|---|---|---|---|---|---|
causallm-14b | 88.26087 | 1.116333 | 705 | 89 | 11 | 805 | community | 1391 |
在 AlpacaEval Leaderboard 胜率 88.26% view raw
DPO 版本的 MT-Behch
Model | MT-Bench |
---|---|
GPT-4 | 8.99 |
GPT-3.5-Turbo | 7.94 |
Zephyr-7b-β (Overfitting) | 7.34 |
Zephyr-7b-α | 6.88 |
CausalLM/14B-DPO-α | 7.618868 |
CausalLM/7B-DPO-α | 7.038125 |
其他语言
我们目前无法为非 QA 任务(英语和中文以外的语言)生成准确的基准模板。 不过,我们将在不久的将来开发其他语言版本的 QA-Task 挑战。
日文基准
Task | Version | Metric | Value | Stderr | |
---|---|---|---|---|---|
jcommonsenseqa-1.1-0.6 | 1.1 | acc | 0.8213 | ± | 0.0115 |
JCommonsenseQA 基准测试结果非常非常接近 Japanese Stable LM Gamma 7B (83.47),当前 SOTA 日文 LM 。然而,我们的模型并未在日文上进行特别的大量文本训练。这似乎能体现元语言的跨语言迁移能力。
🤗 Open LLM 排行榜
- Downloads last month
- 26
Model tree for cgus/CausalLM-14B-exl2
Base model
CausalLM/14B