Upload folder using huggingface_hub
Browse files
README.md
CHANGED
@@ -1,79 +1,91 @@
|
|
1 |
---
|
2 |
-
|
|
|
3 |
tags:
|
4 |
-
- trl
|
5 |
-
- dpo
|
6 |
- DPO
|
7 |
- WeniGPT
|
8 |
-
- generated_from_trainer
|
9 |
base_model: Weni/WeniGPT-Agents-Mistral-1.0.6-SFT-merged
|
10 |
model-index:
|
11 |
-
- name: WeniGPT-Agents-Mistral-1.0.6-SFT-1.0.6-DPO
|
12 |
results: []
|
|
|
13 |
---
|
14 |
|
15 |
-
|
16 |
-
should probably proofread and complete it, then remove this comment. -->
|
17 |
|
18 |
-
|
|
|
19 |
|
20 |
-
This model is a fine-tuned version of [Weni/WeniGPT-Agents-Mistral-1.0.6-SFT-merged](https://huggingface.co/Weni/WeniGPT-Agents-Mistral-1.0.6-SFT-merged) on an unknown dataset.
|
21 |
It achieves the following results on the evaluation set:
|
22 |
-
-
|
23 |
-
- Rewards/chosen: 1.0224
|
24 |
-
- Rewards/rejected: -0.1812
|
25 |
-
- Rewards/accuracies: 0.5
|
26 |
-
- Rewards/margins: 1.2036
|
27 |
-
- Logps/rejected: -55.7310
|
28 |
-
- Logps/chosen: -34.8112
|
29 |
-
- Logits/rejected: -1.8517
|
30 |
-
- Logits/chosen: -1.8185
|
31 |
|
32 |
-
##
|
33 |
|
34 |
-
|
35 |
|
36 |
-
##
|
37 |
|
38 |
-
|
39 |
|
40 |
-
|
|
|
|
|
|
|
|
|
41 |
|
42 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
43 |
|
44 |
-
|
|
|
|
|
|
|
45 |
|
46 |
### Training hyperparameters
|
47 |
|
48 |
The following hyperparameters were used during training:
|
49 |
- learning_rate: 5e-06
|
50 |
-
-
|
51 |
-
-
|
52 |
-
- seed: 42
|
53 |
-
- distributed_type: multi-GPU
|
54 |
-
- num_devices: 4
|
55 |
- gradient_accumulation_steps: 2
|
|
|
56 |
- total_train_batch_size: 16
|
57 |
-
-
|
58 |
-
-
|
59 |
-
-
|
60 |
-
-
|
61 |
-
-
|
62 |
-
- mixed_precision_training: Native AMP
|
63 |
|
64 |
### Training results
|
65 |
|
66 |
-
| Training Loss | Epoch | Step | Validation Loss | Rewards/chosen | Rewards/rejected | Rewards/accuracies | Rewards/margins | Logps/rejected | Logps/chosen | Logits/rejected | Logits/chosen |
|
67 |
-
|:-------------:|:-----:|:----:|:---------------:|:--------------:|:----------------:|:------------------:|:---------------:|:--------------:|:------------:|:---------------:|:-------------:|
|
68 |
-
| 0.5772 | 1.94 | 30 | 0.5171 | 0.5191 | -0.0654 | 0.5 | 0.5845 | -55.3450 | -36.4889 | -1.8456 | -1.8132 |
|
69 |
-
| 0.5125 | 3.87 | 60 | 0.4517 | 0.8981 | -0.1507 | 0.5 | 1.0487 | -55.6293 | -35.2258 | -1.8501 | -1.8170 |
|
70 |
-
| 0.491 | 5.81 | 90 | 0.4369 | 1.0224 | -0.1812 | 0.5 | 1.2036 | -55.7310 | -34.8112 | -1.8517 | -1.8185 |
|
71 |
-
|
72 |
-
|
73 |
### Framework versions
|
74 |
|
75 |
-
-
|
76 |
-
-
|
77 |
-
-
|
78 |
-
-
|
79 |
-
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
+
license: mit
|
3 |
+
library_name: "trl"
|
4 |
tags:
|
|
|
|
|
5 |
- DPO
|
6 |
- WeniGPT
|
|
|
7 |
base_model: Weni/WeniGPT-Agents-Mistral-1.0.6-SFT-merged
|
8 |
model-index:
|
9 |
+
- name: Weni/WeniGPT-Agents-Mistral-1.0.6-SFT-1.0.6-DPO
|
10 |
results: []
|
11 |
+
language: ['pt']
|
12 |
---
|
13 |
|
14 |
+
# Weni/WeniGPT-Agents-Mistral-1.0.6-SFT-1.0.6-DPO
|
|
|
15 |
|
16 |
+
This model is a fine-tuned version of [Weni/WeniGPT-Agents-Mistral-1.0.6-SFT-merged] on the dataset Weni/wenigpt-agent-dpo-1.0.0 with the DPO trainer. It is part of the WeniGPT project for [Weni](https://weni.ai/).
|
17 |
+
Description: Experiment on DPO with other hyperparameters and best SFT model of WeniGPT
|
18 |
|
|
|
19 |
It achieves the following results on the evaluation set:
|
20 |
+
{'eval_loss': 0.4368760883808136, 'eval_runtime': 2.4897, 'eval_samples_per_second': 11.246, 'eval_steps_per_second': 1.607, 'eval_rewards/chosen': 1.0224344730377197, 'eval_rewards/rejected': -0.18119940161705017, 'eval_rewards/accuracies': 0.5, 'eval_rewards/margins': 1.2036339044570923, 'eval_logps/rejected': -55.731048583984375, 'eval_logps/chosen': -34.81117630004883, 'eval_logits/rejected': -1.8517436981201172, 'eval_logits/chosen': -1.8184903860092163, 'epoch': 5.81}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
|
22 |
+
## Intended uses & limitations
|
23 |
|
24 |
+
This model has not been trained to avoid specific intructions.
|
25 |
|
26 |
+
## Training procedure
|
27 |
|
28 |
+
Finetuning was done on the model Weni/WeniGPT-Agents-Mistral-1.0.6-SFT-merged with the following prompt:
|
29 |
|
30 |
+
```
|
31 |
+
---------------------
|
32 |
+
System_prompt:
|
33 |
+
Agora você se chama {name}, você é {occupation} e seu objetivo é {chatbot_goal}. O adjetivo que mais define a sua personalidade é {adjective} e você se comporta da seguinte forma:
|
34 |
+
{instructions_formatted}
|
35 |
|
36 |
+
{context_statement}
|
37 |
+
|
38 |
+
Lista de requisitos:
|
39 |
+
- Responda de forma natural, mas nunca fale sobre um assunto fora do contexto.
|
40 |
+
- Nunca traga informações do seu próprio conhecimento.
|
41 |
+
- Repito é crucial que você responda usando apenas informações do contexto.
|
42 |
+
- Nunca mencione o contexto fornecido.
|
43 |
+
- Nunca mencione a pergunta fornecida.
|
44 |
+
- Gere a resposta mais útil possível para a pergunta usando informações do conexto acima.
|
45 |
+
- Nunca elabore sobre o porque e como você fez a tarefa, apenas responda.
|
46 |
|
47 |
+
|
48 |
+
---------------------
|
49 |
+
|
50 |
+
```
|
51 |
|
52 |
### Training hyperparameters
|
53 |
|
54 |
The following hyperparameters were used during training:
|
55 |
- learning_rate: 5e-06
|
56 |
+
- per_device_train_batch_size: 2
|
57 |
+
- per_device_eval_batch_size: 2
|
|
|
|
|
|
|
58 |
- gradient_accumulation_steps: 2
|
59 |
+
- num_gpus: 4
|
60 |
- total_train_batch_size: 16
|
61 |
+
- optimizer: AdamW
|
62 |
+
- lr_scheduler_type: cosine
|
63 |
+
- num_steps: 90
|
64 |
+
- quantization_type: bitsandbytes
|
65 |
+
- LoRA: ("\n - bits: 4\n - use_exllama: True\n - device_map: auto\n - use_cache: False\n - lora_r: 8\n - lora_alpha: 16\n - lora_dropout: 0.05\n - bias: none\n - target_modules: ['v_proj', 'q_proj']\n - task_type: CAUSAL_LM",)
|
|
|
66 |
|
67 |
### Training results
|
68 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
69 |
### Framework versions
|
70 |
|
71 |
+
- transformers==4.38.2
|
72 |
+
- datasets==2.18.0
|
73 |
+
- peft==0.10.0
|
74 |
+
- safetensors==0.4.2
|
75 |
+
- evaluate==0.4.1
|
76 |
+
- bitsandbytes==0.43
|
77 |
+
- huggingface_hub==0.22.2
|
78 |
+
- seqeval==1.2.2
|
79 |
+
- optimum==1.18.1
|
80 |
+
- auto-gptq==0.7.1
|
81 |
+
- gpustat==1.1.1
|
82 |
+
- deepspeed==0.14.0
|
83 |
+
- wandb==0.16.6
|
84 |
+
- trl==0.8.1
|
85 |
+
- accelerate==0.29.2
|
86 |
+
- coloredlogs==15.0.1
|
87 |
+
- traitlets==5.14.2
|
88 |
+
- autoawq@https://github.com/casper-hansen/AutoAWQ/releases/download/v0.2.4/autoawq-0.2.4+cu118-cp310-cp310-linux_x86_64.whl
|
89 |
+
|
90 |
+
### Hardware
|
91 |
+
- Cloud provided: runpod.io
|