beamaia commited on
Commit
855322a
1 Parent(s): 753c678

Upload folder using huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +73 -42
README.md CHANGED
@@ -1,72 +1,103 @@
1
  ---
2
  license: mit
3
- library_name: peft
4
  tags:
5
- - trl
6
- - kto
7
- - generated_from_trainer
8
  base_model: HuggingFaceH4/zephyr-7b-beta
9
  model-index:
10
- - name: WeniGPT-Agents-Zephyr-1.0.9-KTO
11
  results: []
 
12
  ---
13
 
14
- <!-- This model card has been generated automatically according to the information the Trainer had access to. You
15
- should probably proofread and complete it, then remove this comment. -->
16
 
17
- # WeniGPT-Agents-Zephyr-1.0.9-KTO
 
18
 
19
- This model is a fine-tuned version of [HuggingFaceH4/zephyr-7b-beta](https://huggingface.co/HuggingFaceH4/zephyr-7b-beta) on the None dataset.
20
  It achieves the following results on the evaluation set:
21
- - Loss: 0.5443
22
- - Rewards/chosen: -130.8165
23
- - Logps/chosen: -1605.7635
24
- - Rewards/rejected: -121.0086
25
- - Logps/rejected: -1472.8949
26
- - Kl: 0.0
27
- - Rewards/margins: -16.2952
28
 
29
- ## Model description
30
 
31
- More information needed
32
 
33
- ## Intended uses & limitations
34
 
35
- More information needed
36
 
37
- ## Training and evaluation data
 
 
 
 
38
 
39
- More information needed
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
40
 
41
- ## Training procedure
 
 
 
 
 
 
 
 
42
 
43
  ### Training hyperparameters
44
 
45
  The following hyperparameters were used during training:
46
  - learning_rate: 0.0002
47
- - train_batch_size: 4
48
- - eval_batch_size: 4
49
- - seed: 42
50
  - gradient_accumulation_steps: 4
 
51
  - total_train_batch_size: 16
52
- - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
53
- - lr_scheduler_type: linear
54
- - lr_scheduler_warmup_ratio: 0.03
55
- - training_steps: 145
56
- - mixed_precision_training: Native AMP
57
 
58
  ### Training results
59
 
60
- | Training Loss | Epoch | Step | Validation Loss | Rewards/chosen | Logps/chosen | Rewards/rejected | Logps/rejected | Kl | Rewards/margins |
61
- |:-------------:|:-----:|:----:|:---------------:|:--------------:|:------------:|:----------------:|:--------------:|:---:|:---------------:|
62
- | 0.4555 | 0.34 | 50 | 0.5443 | -148.8861 | -1786.4595 | -135.3632 | -1616.4402 | 0.0 | -20.0430 |
63
- | 0.5714 | 0.68 | 100 | 0.5443 | -130.8165 | -1605.7635 | -121.0086 | -1472.8949 | 0.0 | -16.2952 |
64
-
65
-
66
  ### Framework versions
67
 
68
- - PEFT 0.10.0
69
- - Transformers 4.39.1
70
- - Pytorch 2.1.0+cu118
71
- - Datasets 2.18.0
72
- - Tokenizers 0.15.2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: mit
3
+ library_name: "trl"
4
  tags:
5
+ - KTO
6
+ - WeniGPT
 
7
  base_model: HuggingFaceH4/zephyr-7b-beta
8
  model-index:
9
+ - name: Weni/WeniGPT-Agents-Zephyr-1.0.9-KTO
10
  results: []
11
+ language: ['pt']
12
  ---
13
 
14
+ # Weni/WeniGPT-Agents-Zephyr-1.0.9-KTO
 
15
 
16
+ This model is a fine-tuned version of [HuggingFaceH4/zephyr-7b-beta] on the dataset Weni/wenigpt-agent-1.2.0 with the KTO trainer. It is part of the WeniGPT project for [Weni](https://weni.ai/).
17
+ Description: Hyperparameter search, altering desired and undesired weights for KTO task.
18
 
 
19
  It achieves the following results on the evaluation set:
20
+ {'eval_loss': 0.5443353056907654, 'eval_runtime': 138.5128, 'eval_samples_per_second': 2.166, 'eval_steps_per_second': 0.541, 'eval_rewards/chosen': -130.8164825439453, 'eval_logps/chosen': -1605.7635498046875, 'eval_rewards/rejected': -121.00863647460938, 'eval_logps/rejected': -1472.8948974609375, 'eval_kl': 0.0, 'eval_rewards/margins': -16.295190811157227, 'epoch': 0.99}
 
 
 
 
 
 
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 HuggingFaceH4/zephyr-7b-beta 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
+ Na sua memória você tem esse contexto:
37
+ {context}
38
+
39
+ Lista de requisitos:
40
+ - Responda de forma natural, mas nunca fale sobre um assunto fora do contexto.
41
+ - Nunca traga informações do seu próprio conhecimento.
42
+ - Repito é crucial que você responda usando apenas informações do contexto.
43
+ - Nunca mencione o contexto fornecido.
44
+ - Nunca mencione a pergunta fornecida.
45
+ - Gere a resposta mais útil possível para a pergunta usando informações do conexto acima.
46
+ - Nunca elabore sobre o porque e como você fez a tarefa, apenas responda.
47
+
48
+
49
+ ---------------------
50
+ Question:
51
+ {question}
52
 
53
+
54
+ ---------------------
55
+ Response:
56
+ {answer}
57
+
58
+
59
+ ---------------------
60
+
61
+ ```
62
 
63
  ### Training hyperparameters
64
 
65
  The following hyperparameters were used during training:
66
  - learning_rate: 0.0002
67
+ - per_device_train_batch_size: 4
68
+ - per_device_eval_batch_size: 4
 
69
  - gradient_accumulation_steps: 4
70
+ - num_gpus: 1
71
  - total_train_batch_size: 16
72
+ - optimizer: AdamW
73
+ - lr_scheduler_type: cosine
74
+ - num_steps: 145
75
+ - quantization_type: bitsandbytes
76
+ - LoRA: ("\n - bits: 4\n - use_exllama: True\n - device_map: auto\n - use_cache: False\n - lora_r: 16\n - lora_alpha: 32\n - lora_dropout: 0.05\n - bias: none\n - target_modules: ['q_proj', 'k_proj', 'v_proj', 'o_proj']\n - task_type: CAUSAL_LM",)
77
 
78
  ### Training results
79
 
 
 
 
 
 
 
80
  ### Framework versions
81
 
82
+ - transformers==4.39.1
83
+ - datasets==2.18.0
84
+ - peft==0.10.0
85
+ - safetensors==0.4.2
86
+ - evaluate==0.4.1
87
+ - bitsandbytes==0.43
88
+ - huggingface_hub==0.20.3
89
+ - seqeval==1.2.2
90
+ - optimum==1.17.1
91
+ - auto-gptq==0.7.1
92
+ - gpustat==1.1.1
93
+ - deepspeed==0.14.0
94
+ - wandb==0.16.3
95
+ - # trl==0.8.1
96
+ - git+https://github.com/kawine/trl.git#egg=trl
97
+ - accelerate==0.28.0
98
+ - coloredlogs==15.0.1
99
+ - traitlets==5.14.1
100
+ - autoawq@https://github.com/casper-hansen/AutoAWQ/releases/download/v0.2.0/autoawq-0.2.0+cu118-cp310-cp310-linux_x86_64.whl
101
+
102
+ ### Hardware
103
+ - Cloud provided: runpod.io