ToastyPigeon
commited on
Commit
•
39494e2
1
Parent(s):
7fb4344
Update README.md
Browse files
README.md
CHANGED
@@ -14,146 +14,20 @@ model-index:
|
|
14 |
should probably proofread and complete it, then remove this comment. -->
|
15 |
|
16 |
[<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)
|
17 |
-
<details><summary>See axolotl config</summary>
|
18 |
|
19 |
-
|
20 |
-
```yaml
|
21 |
-
# python -m axolotl.cli.preprocess adventure-l31.yml
|
22 |
-
# accelerate launch -m axolotl.cli.train adventure-l31.yml
|
23 |
-
# python -m axolotl.cli.merge_lora adventure-l31.yml
|
24 |
|
25 |
-
|
26 |
-
model_type: AutoModelForCausalLM
|
27 |
-
tokenizer_type: AutoTokenizer
|
28 |
-
|
29 |
-
load_in_8bit: false
|
30 |
-
load_in_4bit: true
|
31 |
-
strict: false
|
32 |
-
sequence_len: 8192 # 99% vram
|
33 |
-
bf16: auto
|
34 |
-
fp16:
|
35 |
-
tf32: false
|
36 |
-
flash_attention: true
|
37 |
-
special_tokens:
|
38 |
-
|
39 |
-
# Data
|
40 |
-
dataset_prepared_path: last_run_prepared
|
41 |
-
datasets:
|
42 |
-
- path: ColumbidAI/adventure-8k
|
43 |
-
type: completion
|
44 |
-
warmup_steps: 20
|
45 |
-
shuffle_merged_datasets: true
|
46 |
-
|
47 |
-
save_safetensors: true
|
48 |
-
saves_per_epoch: 4
|
49 |
-
save_total_limit: 2
|
50 |
-
|
51 |
-
# WandB
|
52 |
-
wandb_project: L31-A
|
53 |
-
wandb_entity:
|
54 |
-
|
55 |
-
# Iterations
|
56 |
-
num_epochs: 1
|
57 |
-
|
58 |
-
# Output
|
59 |
-
output_dir: ./adventure-command-r-workspace
|
60 |
-
hub_model_id: ToastyPigeon/adventure-nemo-ws
|
61 |
-
hub_strategy: "all_checkpoints"
|
62 |
-
|
63 |
-
# Sampling
|
64 |
-
sample_packing: true
|
65 |
-
pad_to_sequence_len: true
|
66 |
-
|
67 |
-
# Batching
|
68 |
-
gradient_accumulation_steps: 2
|
69 |
-
micro_batch_size: 8
|
70 |
-
gradient_checkpointing: 'unsloth'
|
71 |
-
gradient_checkpointing_kwargs:
|
72 |
-
use_reentrant: true
|
73 |
-
|
74 |
-
#unsloth_cross_entropy_loss: true
|
75 |
-
#unsloth_lora_mlp: true
|
76 |
-
#unsloth_lora_qkv: true
|
77 |
-
#unsloth_lora_o: true
|
78 |
-
|
79 |
-
# Evaluation
|
80 |
-
val_set_size: 0.01
|
81 |
-
evals_per_epoch: 5
|
82 |
-
eval_table_size:
|
83 |
-
eval_max_new_tokens: 256
|
84 |
-
eval_sample_packing: false
|
85 |
-
eval_batch_size: 1
|
86 |
-
|
87 |
-
# LoRA
|
88 |
-
adapter: qlora
|
89 |
-
lora_model_dir:
|
90 |
-
lora_r: 64
|
91 |
-
lora_alpha: 32
|
92 |
-
lora_dropout: 0.125
|
93 |
-
lora_target_linear:
|
94 |
-
lora_fan_in_fan_out:
|
95 |
-
lora_target_modules:
|
96 |
-
- gate_proj
|
97 |
-
- down_proj
|
98 |
-
- up_proj
|
99 |
-
- q_proj
|
100 |
-
- v_proj
|
101 |
-
- k_proj
|
102 |
-
- o_proj
|
103 |
-
lora_modules_to_save:
|
104 |
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
learning_rate: 0.00005
|
109 |
-
lr_scheduler: cosine_with_min_lr
|
110 |
-
lr_scheduler_kwargs:
|
111 |
-
min_lr: 0.000005
|
112 |
-
weight_decay: 0.01
|
113 |
-
max_grad_norm: 20.0
|
114 |
|
115 |
-
|
116 |
-
train_on_inputs: false
|
117 |
-
group_by_length: false
|
118 |
-
early_stopping_patience:
|
119 |
-
local_rank:
|
120 |
-
logging_steps: 1
|
121 |
-
xformers_attention:
|
122 |
-
debug:
|
123 |
-
#deepspeed: /workspace/axolotl/deepspeed_configs/zero3.json # previously blank
|
124 |
-
fsdp:
|
125 |
-
fsdp_config:
|
126 |
|
|
|
127 |
|
128 |
-
|
129 |
-
- axolotl.integrations.liger.LigerPlugin
|
130 |
-
liger_rope: true
|
131 |
-
liger_rms_norm: true
|
132 |
-
liger_swiglu: true
|
133 |
-
liger_fused_linear_cross_entropy: true
|
134 |
-
```
|
135 |
-
|
136 |
-
</details><br>
|
137 |
-
|
138 |
-
# adventure-nemo-ws
|
139 |
-
|
140 |
-
This model is a fine-tuned version of [unsloth/Meta-Llama-3.1-8B](https://huggingface.co/unsloth/Meta-Llama-3.1-8B) on the None dataset.
|
141 |
-
It achieves the following results on the evaluation set:
|
142 |
-
- Loss: 2.3893
|
143 |
-
|
144 |
-
## Model description
|
145 |
-
|
146 |
-
More information needed
|
147 |
-
|
148 |
-
## Intended uses & limitations
|
149 |
-
|
150 |
-
More information needed
|
151 |
-
|
152 |
-
## Training and evaluation data
|
153 |
-
|
154 |
-
More information needed
|
155 |
-
|
156 |
-
## Training procedure
|
157 |
|
158 |
### Training hyperparameters
|
159 |
|
@@ -169,16 +43,6 @@ The following hyperparameters were used during training:
|
|
169 |
- lr_scheduler_warmup_steps: 20
|
170 |
- num_epochs: 1
|
171 |
|
172 |
-
### Training results
|
173 |
-
|
174 |
-
| Training Loss | Epoch | Step | Validation Loss |
|
175 |
-
|:-------------:|:------:|:----:|:---------------:|
|
176 |
-
| 2.2246 | 0.0045 | 1 | 2.4988 |
|
177 |
-
| 2.1034 | 0.2013 | 45 | 2.4257 |
|
178 |
-
| 2.2138 | 0.4027 | 90 | 2.4077 |
|
179 |
-
| 2.1541 | 0.6040 | 135 | 2.3941 |
|
180 |
-
| 2.0555 | 0.8054 | 180 | 2.3893 |
|
181 |
-
|
182 |
|
183 |
### Framework versions
|
184 |
|
|
|
14 |
should probably proofread and complete it, then remove this comment. -->
|
15 |
|
16 |
[<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)
|
|
|
17 |
|
18 |
+
# Meta-Llama-3.1-8B-Adventure-QLoRA
|
|
|
|
|
|
|
|
|
19 |
|
20 |
+
This LoRA is trained on Llama 3.1 8B **base** using completion format.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
|
22 |
+
The datasets used were:
|
23 |
+
- Spring Dragon
|
24 |
+
- Skein
|
|
|
|
|
|
|
|
|
|
|
|
|
25 |
|
26 |
+
This is not an instruct model and **no instruct format was used.**
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
27 |
|
28 |
+
The intended use is with text completion where user input is given with `> User Input`. This is the default for Kobold Lite Adventure mode.
|
29 |
|
30 |
+
If merged into an instruct model, it should impart the flavor of the text adventure data. Use whatever the instruct model's format is for instruct.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
31 |
|
32 |
### Training hyperparameters
|
33 |
|
|
|
43 |
- lr_scheduler_warmup_steps: 20
|
44 |
- num_epochs: 1
|
45 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
46 |
|
47 |
### Framework versions
|
48 |
|