metadata
license: apache-2.0
base_model:
- meta-llama/Llama-3.1-8B-Instruct
A preview version of FuseChat-3.0, under testing...
Training configs
# Model arguments
model_name_or_path: AALF/FuseChat-Llama-3.1-8B-SFT
torch_dtype: null
attn_implementation: flash_attention_2
# Data training arguments
dataset_mixer: FuseChat-Mixture-v3-DPO
dataset_splits:
- train
- test
preprocessing_num_workers: 12
# DPOTrainer arguments
bf16: true
beta: 10
avg_logp: true
gradient_accumulation_steps: 8
gradient_checkpointing: true
gradient_checkpointing_kwargs:
use_reentrant: False
hub_model_id: wrpo-models
learning_rate: 8.0e-7
log_level: info
logging_steps: 5
lr_scheduler_type: cosine
max_length: 2048
max_prompt_length: 1800
num_train_epochs: 1
optim: adamw_torch
output_dir: outputs/FuseChat-Llama-3.1-8B-Instruct
run_name: FuseChat-Llama-3.1-8B-Instruct
per_device_train_batch_size: 2
per_device_eval_batch_size: 4
push_to_hub: false
save_strategy: "steps"
save_steps: 101
save_total_limit: 20
seed: 42
warmup_ratio: 0.1
save_only_model: true
Evaluation Results
| Datasets | Llama3.1-8B-Instruct | FuseChat-Llama-3.1-8B-SFT | FuseChat-Llama-3.1-8B-Instruct |
|---|---|---|---|
| AlpacaEval-2 (LC/WR) | 28.3/28.7 | 41.3/37.7 | 65.4/63.3 |
| Arena-Hard (WR/SC) | 28.1/23.8 | 38.7/29 | 58.2/46.4 |
| MT-Bench | 8.38 | 8.54 | 9 |
| AlignBench v1.1 | 4.61 | 6.25 | 6.69 |
| LiveBench 0831 | 27.6 | 30.2 | 32 |
| GSM8K | 85.9 | 87 | 88 |
| MATH | 50.7 | 54.7 | 55.2 |
| AMC 23 | 25 | 30 | 37.5 |
| MMLU-Pro | 50 | 47.8 | 49.2 |
| MMLU-redux | 67.2 | 68.4 | 69.2 |
| GPQA-Diamond | 33.8 | 37.9 | 34.9 |
| HumanEval | 69.5 | 69.5 | 71.3 |
| MBPP | 75.4 | 71.4 | 72 |
| LiveCodeBench 2408-2411 (all/esay) | 12.3/40.5 | 12.6/39 | 13.1/43.2 |