loubnabnl HF staff commited on
Commit
0691f98
1 Parent(s): 70be2d6

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +29 -64
README.md CHANGED
@@ -1,78 +1,43 @@
1
  ---
2
- base_model: HuggingFaceTB/cosmo2-1.7B-webinst-sc2
3
- tags:
4
- - alignment-handbook
5
- - trl
6
- - dpo
7
- - generated_from_trainer
8
- - trl
9
- - dpo
10
- - generated_from_trainer
11
- datasets:
12
- - HuggingFaceTB/Helpsteer
13
- model-index:
14
- - name: cosmo2-1.7B-webinst-sc2-dpo-helpsteer-ep1
15
- results: []
16
  ---
17
 
18
- <!-- This model card has been generated automatically according to the information the Trainer had access to. You
19
- should probably proofread and complete it, then remove this comment. -->
20
 
21
- [<img src="https://raw.githubusercontent.com/wandb/assets/main/wandb-github-badge-28.svg" alt="Visualize in Weights & Biases" width="200" height="32"/>](https://wandb.ai/loubnabnl/huggingface/runs/ellmeibr)
22
- # cosmo2-1.7B-webinst-sc2-dpo-helpsteer-ep1
23
 
24
- This model is a fine-tuned version of [HuggingFaceTB/cosmo2-1.7B-webinst-sc2](https://huggingface.co/HuggingFaceTB/cosmo2-1.7B-webinst-sc2) on the HuggingFaceTB/Helpsteer dataset.
25
- It achieves the following results on the evaluation set:
26
- - Loss: 0.6672
27
- - Rewards/chosen: -0.0466
28
- - Rewards/rejected: -0.0933
29
- - Rewards/accuracies: 0.5500
30
- - Rewards/margins: 0.0467
31
- - Logps/rejected: -149.4311
32
- - Logps/chosen: -121.9851
33
- - Logits/rejected: 0.8632
34
- - Logits/chosen: 0.9551
35
- - IFEval loose prompt 21.07
36
- - IFEval strict prompt 18.48
37
-
38
- ## Model description
39
 
40
- More information needed
41
 
42
- ## Intended uses & limitations
 
 
 
 
43
 
44
- More information needed
45
 
46
- ## Training and evaluation data
47
 
48
- More information needed
 
49
 
50
- ## Training procedure
51
 
52
- ### Training hyperparameters
53
 
54
- The following hyperparameters were used during training:
55
- - learning_rate: 5e-06
56
- - train_batch_size: 2
57
- - eval_batch_size: 4
58
- - seed: 42
59
- - distributed_type: multi-GPU
60
- - num_devices: 8
61
- - gradient_accumulation_steps: 8
62
- - total_train_batch_size: 128
63
- - total_eval_batch_size: 32
64
- - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
65
- - lr_scheduler_type: cosine
66
- - lr_scheduler_warmup_ratio: 0.1
67
- - num_epochs: 1
68
 
69
- ### Training results
70
-
71
-
72
-
73
- ### Framework versions
74
-
75
- - Transformers 4.42.3
76
- - Pytorch 2.1.2
77
- - Datasets 2.20.0
78
- - Tokenizers 0.19.1
 
1
  ---
2
+ library_name: transformers
3
+ license: apache-2.0
4
+ language:
5
+ - en
 
 
 
 
 
 
 
 
 
 
6
  ---
7
 
 
 
8
 
9
+ # SmolLM-Instruct
 
10
 
11
+ <center>
12
+ <img src="https://huggingface.co/datasets/HuggingFaceTB/images/resolve/main/banner_smol.png" alt="SmolLM" width="1100" height="600">
13
+ </center>
 
 
 
 
 
 
 
 
 
 
 
 
14
 
15
+ ## Table of Contents
16
 
17
+ 1. [Model Summary](##model-summary)
18
+ 2. [Limitations](##limitations)
19
+ 3. [Training](##training)
20
+ 4. [License](##license)
21
+ 5. [Citation](##citation)
22
 
23
+ ## Model Summary
24
 
25
+ SmolLM is a series of state-of-the-art small language models available in three sizes: 135M, 360M, and 1.7B parameters. These models are built on Cosmo-Corpus, a meticulously curated high-quality training dataset. Cosmo-Corpus includes Cosmopedia v2 (28B tokens of synthetic textbooks and stories generated by Mixtral), Python-Edu (4B tokens of educational Python samples from The Stack), and FineWeb-Edu (220B tokens of deduplicated educational web samples from FineWeb). For duther details, we refer to our blogpost TODO.
26
 
27
+ To build SmolLM-Instruct, we instruction tuned the models using publicly available permissive instruction datasets. We trained all three models for one epoch on the permissive subset of the WebInstructSub dataset, combined with StarCoder2-Self-OSS-Instruct. Following this, we performed DPO (Direct Preference Optimization) for one epoch: using HelpSteer for the 135M and 1.7B models, and argilla/dpo-mix-7k for the 360M model. We followed the training parameters from the Zephyr-Gemma recipe in the alignment handbook, but adjusted the SFT (Supervised Fine-Tuning) learning rate to 3e-4.
28
+ [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0)
29
 
30
+ This is the SmolLM-360M-Instruct.
31
 
32
+ # Limitations
33
 
34
+ While SmolLM models have been trained on a diverse dataset including educational content and synthetic texts, they have limitations. The models primarily understand and generate content in English. They can produce text on a variety of topics, but the generated content may not always be factually accurate, logically consistent, or free from biases present in the training data. These models should be used as assistive tools rather than definitive sources of information. Users should always verify important information and critically evaluate any generated content. For a more comprehensive discussion of the models' capabilities and limitations, please refer to our full blog post.
 
 
 
 
 
 
 
 
 
 
 
 
 
35
 
36
+ # Citation
37
+ ```bash
38
+ @misc{allal2024SmolLM,
39
+ title={SmolLM - blazingly fast and remarkably powerful},
40
+ author={Loubna Ben Allal and Anton Lozhkov and Elie Bakouch and Leandro von Werra and Thomas Wolf},
41
+ year={2024},
42
+ }
43
+ ```