|
--- |
|
license: apache-2.0 |
|
--- |
|
# Introduction |
|
CSMPT7b is a large Czech language model continously pretrained for 272b training steps from English [MPT7b](https://huggingface.co/mosaicml/mpt-7b) model. Model was pretrained on ~67b token [Large Czech Collection](https://huggingface.co/datasets/BUT-FIT/but_lcc) using Czech tokenizer, obtained using our vocabulary swap method (see below). |
|
|
|
# Eval |
|
Dev eval at CS-HellaSwag (automatically translated HellaSwag benchmark). |
|
| Model | CS-HellaSwag Accuracy | |
|
|---------------|----------------| |
|
| mistral7b | 0.4992 | |
|
| csmpt@130k steps [released] | __0.5004__ | |
|
| csmpt@100k steps | 0.4959 | |
|
| csmpt@75k steps | 0.4895 | |
|
| csmpt@50k steps | 0.4755 | |
|
| csmpt@26,5k steps | 0.4524 | |
|
|
|
|
|
However, we ran validation over the course of training on CS-Hellaswag, and after 100k steps, the improvements were very noisy if any. |
|
The improvement over mistral7b is not significant. |
|
|
|
<TBD> More evaluation details teaser. |
|
|
|
## Loss |
|
We encountered loss spikes during training. As the model always recovered, and our budget for training 7b model was very constrained, we kept on training. We observed such loss spikes before in our ablations. In these ablations (with GPT-2 small), we found these to be |
|
- (a) influenced by learning rate, the lower the learning rate, less they appear, as it gets higher, they start to appear, and with too high learning rate, the training might diverge on such loss spike. |
|
- (b) in preliminary ablations, they only appear for continuously pretrained models. While we do not know why do they appear, we hypothesize this might be linked to theory on [Adam instability in time-domain correlation of update vectors](https://arxiv.org/pdf/2304.09871.pdf). However |
|
such instabilities were previously observed only for much larger models (larger than 65b). |
|
|
|
### Corpora |
|
The model was trained on 3 corpora, which were hot-swapped during the training. These were collected/filtered during the course of training. |
|
- Corpus #1 was the same we used for our [Czech GPT-2](https://huggingface.co/BUT-FIT/Czech-GPT-2-XL-133k) training (15,621,685,248 tokens). |
|
- Corpus #2 contained 67,981,934,592 tokens, coming mostly from HPLT and CulturaX corpora. |
|
- Corpus #3 is Corpus #2 after we removed proportions of the unappropriate content (which avoided our other checks) through linear classifier. |
|
|
|
|
|
<img src="figures/tloss_full.png" width="900"/> |
|
Figure 1: Training loss. |
|
<img src="figures/tloss_closeup.png" width="900"/> |
|
Figure 2: Training loss closeup. We mark two hotswap places, where the training corpus #1 was switched for internal-corpus #2 and internal-corpus #2.1 respectively. |
|
Additionaly, we perform two ablations: |
|
|
|
- (a) After first hot swap, we continued training on the corpus #1 for a while. |
|
- (b) On step 94,000, the training loss stopped decreasing, increased, and around step 120,000 (near hot swap #2) started decreasing again. To ablate whether this was an effect of hot-swap, |
|
- we resume training from step 93,000 using corpus #3. The optimizer states were reinitialized. |
|
<img src="figures/vloss_closeup.png" width="900"/> |
|
Figure 3: Test loss closeup, testing performed on split of internal-corpus #1. See Figure 2 description for ablation explanation. |
|
|
|
|
|
## Training Method |
|
### Vocabulary Swap |
|
The vocabulary swap was done the same way as our [Czech-GPT-2](https://huggingface.co/BUT-FIT/Czech-GPT-2-XL-133k) model (check it out for comprehensive description.) |
|
We managed to align 4,177 english tokens with corresponding czech tokens. |
|
|
|
## Hyperparameters |
|
Not mentioned hyperparameters were kept the same as for MPT. |
|
| **Name** | **Value** | **Note** | |
|
|----------------------------|---------------|----------------------------------------------------------------------------------------------| |
|
| training sw | llm-foundry | We've done some minor patching (e.g., to allow DDP sync over file) | |
|
| dataset_type | Concat | Sequences at the model's input were concatenated up to `$max_seq_len`, divided by EOS token. | |
|
| tokenizer_size | 64k | Same as in [Czech-GPT-2](https://huggingface.co/BUT-FIT/Czech-GPT-2-XL-133k) | |
|
| max_seq_len | 2048 | | |
|
| batch_size | 1024 | | |
|
| learning_rate | 1.0e-4 | | |
|
| optimizer | LionW | | |
|
| optimizer_betas | 0.9/0.95 | | |
|
| optimizer_weight_decay | 0 | | |
|
| optimizer_eps | 1.0e-08 | | |
|
| gradient_clipping_max_norm | 1.0 | | |
|
| attn_impl | flash2 | we used triton flash-attn 1 implementation for initial ~60k steps | |
|
| positional_encoding | alibi | | |
|
| fsdp | FULL_SHARD | (we had implementation issues with hybrid sharding in llm-foundry) | |
|
| precision | bf16 | | |
|
| scheduler | cosine | | |
|
| scheduler_warmup | 100 steps | | |
|
| scheduler_steps | 170,000 | | |
|
| scheduler_alpha | 0.1 | So LR on last step is 0.1*(vanilla LR) | |
|
|
|
|
|
# Usage |
|
## How to Setup Environment |
|
```bash |
|
pip install transformers==4.37.2 torch==2.1.2 einops==0.7.0 |
|
|
|
# be sure to install right flash-attn, we use torch compiled with CUDA 12.1, no ABI, python 3.9, Linux x86_64 architecture |
|
pip install https://github.com/Dao-AILab/flash-attention/releases/download/v2.5.3/flash_attn-2.5.3+cu122torch2.1cxx11abiFALSE-cp39-cp39-linux_x86_64.whl |
|
``` |
|
|
|
## Running the Code |
|
```python |
|
import torch |
|
import transformers |
|
from transformers import pipeline |
|
|
|
name = 'BUT-FIT/csmpt7b' |
|
|
|
config = transformers.AutoConfig.from_pretrained(name, trust_remote_code=True) |
|
config.init_device = 'cuda:0' # For fast initialization directly on GPU! |
|
model = transformers.AutoModelForCausalLM.from_pretrained( |
|
name, |
|
config=config, |
|
torch_dtype=torch.bfloat16, # Load model weights in bfloat16 |
|
trust_remote_code=True |
|
) |
|
|
|
tokenizer = transformers.AutoTokenizer.from_pretrained(name, trust_remote_code=True) |
|
|
|
pipe = pipeline('text-generation', model=model, tokenizer=tokenizer, device='cuda:0') |
|
|
|
with torch.autocast('cuda', dtype=torch.bfloat16): |
|
print( |
|
pipe('Nejznámějším českým spisovatelem ', |
|
max_new_tokens=100, |
|
top_p=0.95, |
|
repetition_penalty=1.0, |
|
do_sample=True, |
|
use_cache=True)) |
|
|
|
``` |
|
# Training Data |
|
We release most (95.79%) of our training data corpus as [BUT-Large Czech Collection](https://huggingface.co/datasets/BUT-FIT/but_lcc). |
|
|
|
|
|
# Our Release Plan |
|
| Stage | Description | Date | |
|
|---------------|----------------|----------------| |
|
| 1 | 'Best' model + training data | 11.03.2024 |
|
| 2 | All checkpoints + training code| |
|
| 3 | __Benczechmark__ a collection of Czech datasets for few-shot LLM evaluation **Get in touch if you want to contribute!** | |
|
| 4 | Preprint Publication | |
|
|
|
## Getting in Touch |
|
For further questions, email to `[email protected]`. |
|
|
|
# Disclaimer |
|
This is a probabilistic model, it can output stochastic information. Authors are not responsible for the model outputs. Use at your own risk. |
|
|
|
# Acknowledgement |
|
This work was supported by NAKI III program of Ministry of Culture Czech Republic, project semANT --- |
|
"Sémantický průzkumník textového kulturního dědictví" grant no. `DH23P03OVV060` and |
|
by the Ministry of Education, Youth and Sports of the Czech Republic through the e-INFRA CZ (ID:`90254`). |
|
|
|
# Citation |
|
```bibtex |
|
@article{benczechmark, |
|
author = {Martin Fajčík, Martin Dočekal, Jan Doležal, Karel Beneš, Michal Hradiš}, |
|
title = {BenCzechMark: Machine Language Understanding Benchmark for Czech Language}, |
|
journal = {arXiv preprint arXiv:insert-arxiv-number-here}, |
|
year = {2024}, |
|
month = {March}, |
|
eprint = {insert-arxiv-number-here}, |
|
archivePrefix = {arXiv}, |
|
primaryClass = {cs.CL}, |
|
} |
|
|
|
``` |
