Update README.md
Browse files
README.md
CHANGED
@@ -35,8 +35,9 @@ Awaiting results...
|
|
35 |
|
36 |
## Training Configuration
|
37 |
|
38 |
-
<details><summary>Click here for Axolotl
|
39 |
|
|
|
40 |
```yaml
|
41 |
base_model: arcee-ai/Llama-3.1-SuperNova-Lite
|
42 |
model_type: AutoModelForCausalLM
|
@@ -159,6 +160,100 @@ deepspeed:
|
|
159 |
weight_decay: 0.05
|
160 |
fsdp:
|
161 |
fsdp_config:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
162 |
special_tokens:
|
163 |
pad_token: <|finetune_right_pad_id|>
|
164 |
eos_token: <|eot_id|>
|
|
|
35 |
|
36 |
## Training Configuration
|
37 |
|
38 |
+
<details><summary>Click here for Axolotl configs</summary>
|
39 |
|
40 |
+
SFT
|
41 |
```yaml
|
42 |
base_model: arcee-ai/Llama-3.1-SuperNova-Lite
|
43 |
model_type: AutoModelForCausalLM
|
|
|
160 |
weight_decay: 0.05
|
161 |
fsdp:
|
162 |
fsdp_config:
|
163 |
+
special_tokens:
|
164 |
+
pad_token: <|finetune_right_pad_id|>
|
165 |
+
eos_token: <|eot_id|>
|
166 |
+
```
|
167 |
+
|
168 |
+
KTO
|
169 |
+
```
|
170 |
+
base_model: jeiku/Aura-8B
|
171 |
+
model_type: AutoModelForCausalLM
|
172 |
+
tokenizer_type: AutoTokenizer
|
173 |
+
|
174 |
+
load_in_8bit: false
|
175 |
+
load_in_4bit: false
|
176 |
+
strict: false
|
177 |
+
|
178 |
+
hub_model_id: jeiku/aurakto
|
179 |
+
hub_strategy: "all_checkpoints"
|
180 |
+
push_dataset_to_hub:
|
181 |
+
hf_use_auth_token: true
|
182 |
+
|
183 |
+
chat_template: llama3
|
184 |
+
|
185 |
+
rl: kto
|
186 |
+
rl_beta: 0.2
|
187 |
+
kto_desirable_weight: 0.2
|
188 |
+
|
189 |
+
datasets:
|
190 |
+
- path: anthracite-core/full-opus-chosen-hermes-rejected-kto-v1
|
191 |
+
type: llama3.argilla
|
192 |
+
|
193 |
+
shuffle_merged_datasets: true
|
194 |
+
val_set_size: 0.0
|
195 |
+
output_dir: ./outputs/out
|
196 |
+
|
197 |
+
adapter: lora
|
198 |
+
lora_model_dir:
|
199 |
+
|
200 |
+
lora_r: 32
|
201 |
+
lora_alpha: 64
|
202 |
+
lora_dropout: 0.05
|
203 |
+
lora_target_linear: true
|
204 |
+
lora_fan_in_fan_out:
|
205 |
+
|
206 |
+
sequence_len: 8192
|
207 |
+
sample_packing: false
|
208 |
+
eval_sample_packing: false
|
209 |
+
pad_to_sequence_len: false
|
210 |
+
|
211 |
+
wandb_project: Aura-8B
|
212 |
+
wandb_entity:
|
213 |
+
wandb_watch:
|
214 |
+
wandb_name: Aura-8B
|
215 |
+
wandb_log_model:
|
216 |
+
|
217 |
+
gradient_accumulation_steps: 16
|
218 |
+
micro_batch_size: 2
|
219 |
+
num_epochs: 2
|
220 |
+
max_steps: 500
|
221 |
+
|
222 |
+
optimizer: adamw_8bit
|
223 |
+
lr_scheduler: cosine
|
224 |
+
learning_rate: 0.0001
|
225 |
+
weight_decay: 0.05
|
226 |
+
|
227 |
+
train_on_inputs: false
|
228 |
+
group_by_length: false
|
229 |
+
bf16: auto
|
230 |
+
fp16:
|
231 |
+
tf32: true
|
232 |
+
|
233 |
+
gradient_checkpointing: true
|
234 |
+
gradient_checkpointing_kwargs:
|
235 |
+
use_reentrant: true
|
236 |
+
remove_unused_columns: false
|
237 |
+
early_stopping_patience:
|
238 |
+
resume_from_checkpoint:
|
239 |
+
local_rank:
|
240 |
+
logging_steps: 1
|
241 |
+
xformers_attention:
|
242 |
+
flash_attention: true
|
243 |
+
|
244 |
+
warmup_steps: 10
|
245 |
+
evals_per_epoch: 2
|
246 |
+
eval_table_size:
|
247 |
+
eval_max_new_tokens:
|
248 |
+
saves_per_epoch: 1
|
249 |
+
|
250 |
+
debug:
|
251 |
+
deepspeed:
|
252 |
+
fsdp:
|
253 |
+
fsdp_config:
|
254 |
+
fsdp:
|
255 |
+
fsdp_config:
|
256 |
+
|
257 |
special_tokens:
|
258 |
pad_token: <|finetune_right_pad_id|>
|
259 |
eos_token: <|eot_id|>
|