Update README.md
Browse files
README.md
CHANGED
|
@@ -4,4 +4,46 @@ base_model:
|
|
| 4 |
- meta-llama/Llama-3.1-8B-Instruct
|
| 5 |
---
|
| 6 |
|
| 7 |
-
A preview version of FuseChat-3.0, under testing...
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 4 |
- meta-llama/Llama-3.1-8B-Instruct
|
| 5 |
---
|
| 6 |
|
| 7 |
+
A preview version of FuseChat-3.0, under testing...
|
| 8 |
+
Training configs:
|
| 9 |
+
# Model arguments
|
| 10 |
+
model_name_or_path: AALF/FuseChat-Llama-3.1-8B-SFT
|
| 11 |
+
torch_dtype: null
|
| 12 |
+
attn_implementation: flash_attention_2
|
| 13 |
+
|
| 14 |
+
|
| 15 |
+
# Data training arguments
|
| 16 |
+
dataset_mixer: FuseChat-Mixture-v3
|
| 17 |
+
dataset_splits:
|
| 18 |
+
- train
|
| 19 |
+
- test
|
| 20 |
+
preprocessing_num_workers: 12
|
| 21 |
+
|
| 22 |
+
# DPOTrainer arguments
|
| 23 |
+
bf16: true
|
| 24 |
+
beta: 10
|
| 25 |
+
avg_logp: true
|
| 26 |
+
gradient_accumulation_steps: 8
|
| 27 |
+
gradient_checkpointing: true
|
| 28 |
+
gradient_checkpointing_kwargs:
|
| 29 |
+
use_reentrant: False
|
| 30 |
+
hub_model_id: wrpo-models
|
| 31 |
+
learning_rate: 8.0e-7
|
| 32 |
+
log_level: info
|
| 33 |
+
logging_steps: 5
|
| 34 |
+
lr_scheduler_type: cosine
|
| 35 |
+
max_length: 2048
|
| 36 |
+
max_prompt_length: 1800
|
| 37 |
+
num_train_epochs: 1
|
| 38 |
+
optim: adamw_torch
|
| 39 |
+
output_dir: outputs/FuseChat-Llama-3.1-8B-Instruct
|
| 40 |
+
run_name: FuseChat-Llama-3.1-8B-Instruct
|
| 41 |
+
per_device_train_batch_size: 2
|
| 42 |
+
per_device_eval_batch_size: 4
|
| 43 |
+
push_to_hub: false
|
| 44 |
+
save_strategy: "steps"
|
| 45 |
+
save_steps: 101
|
| 46 |
+
save_total_limit: 20
|
| 47 |
+
seed: 42
|
| 48 |
+
warmup_ratio: 0.1
|
| 49 |
+
save_only_model: true
|