Improve language tag
#1
by
lbourdois - opened
README.md
CHANGED
|
@@ -1,174 +1,185 @@
|
|
| 1 |
-
|
| 2 |
-
|
| 3 |
-
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
-
|
| 9 |
-
|
| 10 |
-
-
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
](https://hf.co/QuantFactory)
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
# QuantFactory/Teleut-7b-GGUF
|
| 27 |
+
This is quantized version of [allura-org/Teleut-7b](https://huggingface.co/allura-org/Teleut-7b) created using llama.cpp
|
| 28 |
+
|
| 29 |
+
# Original Model Card
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
# Teleut 7b
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
|
| 36 |
+
A replication attempt of Tulu 3 on the Qwen 2.5 base models.
|
| 37 |
+
|
| 38 |
+
## Evals (so far)
|
| 39 |
+
| | Teleut 7B (measured) | Tülu 3 SFT 8B (reported) | Qwen 2.5 7B Instruct (reported) | Ministral 8B (reported) | Mistral 7B v0.3 (reported)
|
| 40 |
+
|-------------------------|----------------------|--------------------------|---------------------------------|-------------------------|---------------------------
|
| 41 |
+
|BBH (3 shot, CoT) |*64.4%* |**67.9%** |21.7% |56.2% |47.0%<sup>NLL</sup>
|
| 42 |
+
|GSM8K (8 shot, CoT) |78.5% |76.2% |**83.8%** |*80.0%* |xx.x%
|
| 43 |
+
|IFEval (prompt loose) |66.3% |*72.8%* |**74.7%** |56.4% |53.0%
|
| 44 |
+
|MMLU (0 shot, CoT) |*73.2%* |65.9% |**76.6%** |68.5% |30.7%<sup>5-shot</sup>
|
| 45 |
+
|MMLU Pro (0 shot, CoT) |*48.3%* |44.3% |**56.3%**<sup>Unknown</sup> |32.9%<sup>5-shot</sup> |30.7%<sup>5-shot</sup>
|
| 46 |
+
|PopQA (15 shot) |18.9% |**29.3%** |18.1% |*20.2%* |xx.x%
|
| 47 |
+
|TruthfulQA |47.2% |46.8% |**63.1%** |*55.5%* |xx.x%
|
| 48 |
+
|
| 49 |
+
## Credits
|
| 50 |
+
Big thanks to Retis Labs for being providing my 8xH100 polycule used to train and test this model!
|
| 51 |
+
Another big thanks to AllenAI for publishing the Tülu 3 data and model series (as well as the paper and details on training), as well as Alibaba for training the original Qwen 2.5 base model series!
|
| 52 |
+
|
| 53 |
+
```
|
| 54 |
+
@article{lambert2024tulu3,
|
| 55 |
+
title = {Tülu 3: Pushing Frontiers in Open Language Model Post-Training},
|
| 56 |
+
author = {
|
| 57 |
+
Nathan Lambert and
|
| 58 |
+
Jacob Morrison and
|
| 59 |
+
Valentina Pyatkin and
|
| 60 |
+
Shengyi Huang and
|
| 61 |
+
Hamish Ivison and
|
| 62 |
+
Faeze Brahman and
|
| 63 |
+
Lester James V. Miranda and
|
| 64 |
+
Alisa Liu and
|
| 65 |
+
Nouha Dziri and
|
| 66 |
+
Shane Lyu and
|
| 67 |
+
Yuling Gu and
|
| 68 |
+
Saumya Malik and
|
| 69 |
+
Victoria Graf and
|
| 70 |
+
Jena D. Hwang and
|
| 71 |
+
Jiangjiang Yang and
|
| 72 |
+
Ronan Le Bras and
|
| 73 |
+
Oyvind Tafjord and
|
| 74 |
+
Chris Wilhelm and
|
| 75 |
+
Luca Soldaini and
|
| 76 |
+
Noah A. Smith and
|
| 77 |
+
Yizhong Wang and
|
| 78 |
+
Pradeep Dasigi and
|
| 79 |
+
Hannaneh Hajishirzi
|
| 80 |
+
},
|
| 81 |
+
year = {2024},
|
| 82 |
+
email = {tulu@allenai.org}
|
| 83 |
+
}
|
| 84 |
+
```
|
| 85 |
+
|
| 86 |
+
## Training procedure
|
| 87 |
+
|
| 88 |
+
[<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/axolotl-ai-cloud/axolotl)
|
| 89 |
+
|
| 90 |
+
### Training hyperparameters
|
| 91 |
+
|
| 92 |
+
The following hyperparameters were used during training:
|
| 93 |
+
- learning_rate: 3.5e-06
|
| 94 |
+
- train_batch_size: 8
|
| 95 |
+
- eval_batch_size: 8
|
| 96 |
+
- seed: 42
|
| 97 |
+
- distributed_type: multi-GPU
|
| 98 |
+
- num_devices: 8
|
| 99 |
+
- gradient_accumulation_steps: 2
|
| 100 |
+
- total_train_batch_size: 128
|
| 101 |
+
- total_eval_batch_size: 64
|
| 102 |
+
- optimizer: Use paged_ademamix_8bit and the args are:
|
| 103 |
+
No additional optimizer arguments
|
| 104 |
+
- lr_scheduler_type: cosine
|
| 105 |
+
- lr_scheduler_warmup_steps: 370
|
| 106 |
+
- num_epochs: 1
|
| 107 |
+
|
| 108 |
+
### Framework versions
|
| 109 |
+
|
| 110 |
+
- Transformers 4.46.3
|
| 111 |
+
- Pytorch 2.5.1+cu124
|
| 112 |
+
- Datasets 3.1.0
|
| 113 |
+
- Tokenizers 0.20.3
|
| 114 |
+
|
| 115 |
+
### Configuration
|
| 116 |
+
<details><summary>See axolotl config</summary>
|
| 117 |
+
|
| 118 |
+
axolotl version: `0.5.2`
|
| 119 |
+
```yaml
|
| 120 |
+
base_model: Qwen/Qwen2.5-7B
|
| 121 |
+
|
| 122 |
+
plugins:
|
| 123 |
+
- axolotl.integrations.liger.LigerPlugin
|
| 124 |
+
liger_rope: true
|
| 125 |
+
liger_rms_norm: true
|
| 126 |
+
liger_glu_activation: true
|
| 127 |
+
liger_fused_linear_cross_entropy: true
|
| 128 |
+
|
| 129 |
+
strict: false
|
| 130 |
+
|
| 131 |
+
chat_template: chatml
|
| 132 |
+
datasets:
|
| 133 |
+
- path: allenai/tulu-3-sft-mixture
|
| 134 |
+
type: chat_template
|
| 135 |
+
split: train
|
| 136 |
+
field_messages: messages
|
| 137 |
+
|
| 138 |
+
dataset_prepared_path: last_run_prepared
|
| 139 |
+
#val_set_size: 0.02
|
| 140 |
+
output_dir: ./ckpts
|
| 141 |
+
|
| 142 |
+
sequence_len: 8192
|
| 143 |
+
#sample_packing: true
|
| 144 |
+
pad_to_sequence_len: true
|
| 145 |
+
|
| 146 |
+
wandb_project: qwen-2.5-7b-sft
|
| 147 |
+
wandb_entity:
|
| 148 |
+
wandb_watch:
|
| 149 |
+
wandb_name:
|
| 150 |
+
wandb_log_model:
|
| 151 |
+
|
| 152 |
+
gradient_accumulation_steps: 2
|
| 153 |
+
micro_batch_size: 8
|
| 154 |
+
num_epochs: 1
|
| 155 |
+
optimizer: paged_ademamix_8bit
|
| 156 |
+
lr_scheduler: cosine
|
| 157 |
+
learning_rate: 3.5e-6
|
| 158 |
+
|
| 159 |
+
train_on_inputs: false
|
| 160 |
+
group_by_length: false
|
| 161 |
+
bf16: auto
|
| 162 |
+
fp16:
|
| 163 |
+
tf32: false
|
| 164 |
+
|
| 165 |
+
gradient_checkpointing: true
|
| 166 |
+
gradient_checkpointing_kwargs:
|
| 167 |
+
use_reentrant: false
|
| 168 |
+
early_stopping_patience:
|
| 169 |
+
resume_from_checkpoint:
|
| 170 |
+
logging_steps: 1
|
| 171 |
+
xformers_attention:
|
| 172 |
+
flash_attention: true
|
| 173 |
+
|
| 174 |
+
deepspeed: deepspeed_configs/zero3_bf16.json
|
| 175 |
+
|
| 176 |
+
warmup_steps: 370
|
| 177 |
+
#evals_per_epoch: 4
|
| 178 |
+
eval_table_size:
|
| 179 |
+
saves_per_epoch: 2
|
| 180 |
+
debug:
|
| 181 |
+
weight_decay: 0.0
|
| 182 |
+
|
| 183 |
+
```
|
| 184 |
+
|
| 185 |
+
</details><br>
|