AlphaMonarch-laser

AlphaMonarch-laser is a DPO fine-tuned of mlabonne/NeuralMonarch-7B using the argilla/OpenHermes2.5-dpo-binarized-alpha preference dataset but achieves better performance then mlabonne/AlphaMonarch-7B using LaserQLoRA. I have fine-tuned this model only on half of the projections, but have achieved better results as compared to the version released by Maximme Labonne. I have trained this model for 1080 steps.

AlphaMonarch-laser is ranking 1 on YALL - Yet Another LLM Leaderboard.

🏆 Evaluation results

Nous Benchmark

AGIEVAL

Task	Version	Metric	Value	StdErr
agieval_aqua_rat	0	acc	28.35%	2.83%
agieval_aqua_rat	0	acc_norm	26.38%	2.77%
agieval_logiqa_en	0	acc	38.25%	1.91%
agieval_logiqa_en	0	acc_norm	38.10%	1.90%
agieval_lsat_ar	0	acc	23.91%	2.82%
agieval_lsat_ar	0	acc_norm	23.48%	2.80%
agieval_lsat_lr	0	acc	52.75%	2.21%
agieval_lsat_lr	0	acc_norm	53.92%	2.21%
agieval_lsat_rc	0	acc	66.91%	2.87%
agieval_lsat_rc	0	acc_norm	67.29%	2.87%
agieval_sat_en	0	acc	78.64%	2.86%
agieval_sat_en	0	acc_norm	78.64%	2.86%
agieval_sat_en_without_passage	0	acc	45.15%	3.48%
agieval_sat_en_without_passage	0	acc_norm	44.17%	3.47%
agieval_sat_math	0	acc	33.18%	3.18%
agieval_sat_math	0	acc_norm	31.36%	3.14%
Average: 28.41%

GPT4ALL

Task	Version	Metric	Value	StdErr
arc_challenge	0	acc	66.30%	± 1.38%
		acc_norm	68.26%	± 1.36%
arc_easy	0	acc	86.57%	± 0.70%
		acc_norm	80.81%	± 0.81%
boolq	1	acc	87.16%	± 0.59%
hellaswag	0	acc	69.60%	± 0.46%
		acc_norm	87.45%	± 0.33%
openbookqa	0	acc	39.20%	± 2.19%
		acc_norm	49.60%	± 2.24%
piqa	0	acc	83.03%	± 0.88%
		acc_norm	84.87%	± 0.84%
winogrande	0	acc	81.06%	± 1.10%
Average: 76.98%

TRUTHFUL-QA

Task	Version	Metric	Value	StdErr
truthfulqa_mc	1	mc1	63.04%	± 1.69%
truthfulqa_mc	1	mc2	78.39%	± 1.37%
Average: 70.71%

BIGBENCH

Task	Version	Metric	Value	StdErr
bigbench_causal_judgement	0	multiple_choice_grade	60.00%	± 3.56%
bigbench_date_understanding	0	multiple_choice_grade	62.06%	± 2.53%
bigbench_disambiguation_qa	0	multiple_choice_grade	54.26%	± 3.11%
bigbench_geometric_shapes	0	multiple_choice_grade	23.96%	± 2.26%
		exact_str_match	0.00%	± 0.00%
bigbench_logical_deduction_five_objects	0	multiple_choice_grade	32.80%	± 2.10%
bigbench_logical_deduction_seven_objects	0	multiple_choice_grade	23.86%	± 1.61%
bigbench_logical_deduction_three_objects	0	multiple_choice_grade	59.33%	± 2.84%
bigbench_movie_recommendation	0	multiple_choice_grade	58.00%	± 2.21%
bigbench_navigate	0	multiple_choice_grade	56.00%	± 1.57%
bigbench_reasoning_about_colored_objects	0	multiple_choice_grade	69.20%	± 1.03%
bigbench_ruin_names	0	multiple_choice_grade	55.36%	± 2.35%
bigbench_salient_translation_error_detection	0	multiple_choice_grade	41.48%	± 1.56%
bigbench_snarks	0	multiple_choice_grade	73.48%	± 3.29%
bigbench_sports_understanding	0	multiple_choice_grade	76.06%	± 1.36%
bigbench_temporal_sequences	0	multiple_choice_grade	55.50%	± 1.57%
bigbench_tracking_shuffled_objects_five_objects	0	multiple_choice_grade	23.28%	± 1.20%
bigbench_tracking_shuffled_objects_seven_objects	0	multiple_choice_grade	19.37%	± 0.94%
bigbench_tracking_shuffled_objects_three_objects	0	multiple_choice_grade	59.33%	± 2.84%
Average: 55.37%

Openllm Benchmark

Task	Version	Metric	Value		Stderr
arc_challenge	0	acc	70.12	±	1.30
		acc_norm	73.27	±	1.29
hellaswag	0	acc	71.80	±	0.44
		acc_norm	89.20	±	0.30
gsm8k	0	acc	66.77	±	1.2
winogrande	0	acc	84.6	±	1.0

Average: 73.5%

TruthfulQA

Task	Version	Metric	Value		Stderr
truthfulqa_mc	1	mc1	62.79	±	1.69
		mc2	77.90	±	1.37

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-07
• train_batch_size: 1
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 100
• training_steps: 1080

📝 Axolotl Configuration

base_model: mlabonne/NeuralMonarch-7B
model_type: MistralForCausalLM
tokenizer_type: LlamaTokenizer
is_mistral_derived_model: true
load_in_8bit: false
load_in_4bit: true
strict: false
rl: dpo
chat_template: chatml
datasets:
  - path: mlabonne/chatml-OpenHermes2.5-dpo-binarized-alpha
    split: train
    type: chatml.intel
dataset_prepared_path:
val_set_size: 0.01
output_dir: ./out
adapter: qlora
lora_model_dir:
sequence_len: 1800
sample_packing: false
pad_to_sequence_len: false
lora_r: 16
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
lora_target_modules:
 - layers.1.self_attn.q_proj
 - layers.0.self_attn.q_proj
 - layers.15.self_attn.q_proj
 - layers.12.self_attn.q_proj
 - layers.11.self_attn.q_proj
 - layers.14.self_attn.q_proj
 - layers.9.self_attn.q_proj
 - layers.16.self_attn.q_proj
 - layers.30.self_attn.q_proj
 - layers.18.self_attn.q_proj
 - layers.13.self_attn.q_proj
 - layers.10.self_attn.q_proj
 - layers.7.self_attn.q_proj
 - layers.8.self_attn.q_proj
 - layers.4.self_attn.q_proj
 - layers.19.self_attn.q_proj
 - layers.27.self_attn.k_proj
 - layers.24.self_attn.k_proj
 - layers.25.self_attn.k_proj
 - layers.22.self_attn.k_proj
 - layers.26.self_attn.k_proj
 - layers.29.self_attn.k_proj
 - layers.23.self_attn.k_proj
 - layers.28.self_attn.k_proj
 - layers.21.self_attn.k_proj
 - layers.31.self_attn.k_proj
 - layers.30.self_attn.k_proj
 - layers.20.self_attn.k_proj
 - layers.5.self_attn.k_proj
 - layers.19.self_attn.k_proj
 - layers.17.self_attn.k_proj
 - layers.18.self_attn.k_proj
 - layers.19.self_attn.v_proj
 - layers.24.self_attn.v_proj
 - layers.18.self_attn.v_proj
 - layers.5.self_attn.v_proj
 - layers.3.self_attn.v_proj
 - layers.16.self_attn.v_proj
 - layers.23.self_attn.v_proj
 - layers.27.self_attn.v_proj
 - layers.25.self_attn.v_proj
 - layers.26.self_attn.v_proj
 - layers.20.self_attn.v_proj
 - layers.6.self_attn.v_proj
 - layers.15.self_attn.v_proj
 - layers.17.self_attn.v_proj
 - layers.29.self_attn.v_proj
 - layers.22.self_attn.v_proj
 - layers.12.self_attn.o_proj
 - layers.9.self_attn.o_proj
 - layers.14.self_attn.o_proj
 - layers.0.self_attn.o_proj
 - layers.6.self_attn.o_proj
 - layers.8.self_attn.o_proj
 - layers.10.self_attn.o_proj
 - layers.11.self_attn.o_proj
 - layers.13.self_attn.o_proj
 - layers.24.self_attn.o_proj
 - layers.7.self_attn.o_proj
 - layers.15.self_attn.o_proj
 - layers.5.self_attn.o_proj
 - layers.17.self_attn.o_proj
 - layers.25.self_attn.o_proj
 - layers.4.self_attn.o_proj
 - layers.31.mlp.gate_proj
 - layers.30.mlp.gate_proj
 - layers.4.mlp.gate_proj
 - layers.3.mlp.gate_proj
 - layers.29.mlp.gate_proj
 - layers.28.mlp.gate_proj
 - layers.6.mlp.gate_proj
 - layers.27.mlp.gate_proj
 - layers.5.mlp.gate_proj
 - layers.26.mlp.gate_proj
 - layers.25.mlp.gate_proj
 - layers.7.mlp.gate_proj
 - layers.2.mlp.gate_proj
 - layers.24.mlp.gate_proj
 - layers.23.mlp.gate_proj
 - layers.10.mlp.gate_proj
 - layers.6.mlp.up_proj
 - layers.4.mlp.up_proj
 - layers.5.mlp.up_proj
 - layers.27.mlp.up_proj
 - layers.25.mlp.up_proj
 - layers.26.mlp.up_proj
 - layers.17.mlp.up_proj
 - layers.24.mlp.up_proj
 - layers.7.mlp.up_proj
 - layers.10.mlp.up_proj
 - layers.3.mlp.up_proj
 - layers.11.mlp.up_proj
 - layers.23.mlp.up_proj
 - layers.9.mlp.up_proj
 - layers.14.mlp.up_proj
 - layers.18.mlp.up_proj
 - layers.19.mlp.down_proj
 - layers.20.mlp.down_proj
 - layers.18.mlp.down_proj
 - layers.21.mlp.down_proj
 - layers.29.mlp.down_proj
 - layers.1.mlp.down_proj
 - layers.22.mlp.down_proj
 - layers.28.mlp.down_proj
 - layers.23.mlp.down_proj
 - layers.30.mlp.down_proj
 - layers.17.mlp.down_proj
 - layers.4.mlp.down_proj
 - layers.2.mlp.down_proj
 - layers.15.mlp.down_proj
 - layers.5.mlp.down_proj
wandb_project: axolotl
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:
gradient_accumulation_steps: 8
micro_batch_size: 1
num_epochs: 1
optimizer: paged_adamw_32bit
lr_scheduler: cosine
learning_rate: 5e-7
train_on_inputs: false
group_by_length: false
bf16: true
fp16: false
tf32: true
gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true
warmup_steps: 100
evals_per_epoch: 1
eval_table_size:
eval_table_max_new_tokens: 128
save_steps: 1080
max_steps: 1080
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:

Framework versions

• Transformers 4.38.0.dev0
• Pytorch 2.1.2+cu118
• Datasets 2.17.0
• Tokenizers 0.15.0
• axolotl: 0.4.0