Views
No views yet

<|im_start|>system
{system message}<|im_end|>
<|im_start|>user
{user message}<|im_end|>
<|im_start|>assistant
{response}<|im_end|>Temperature - 0.8
Min-P - 0.05
Top-A - 0.3
Repetition Penalty - 1.030.4.11# Model
2base_model: arcee-ai/SuperNova-Medius
3strict: false
4
5# Liger Kernels (optimization)
6plugins:
7 - axolotl.integrations.liger.LigerPlugin
8liger_rope: true
9liger_rms_norm: true
10liger_swiglu: true
11liger_fused_linear_cross_entropy: true
12
13# Output and HuggingFace
14output_dir: /home/kearm/axolotl/TQ-2.5-14B-Sugarquill
15hub_model_id: allura-org/TQ-2.5-14B-Sugarquill-LoRA
16hf_use_auth_token: true
17hub_strategy: "all_checkpoints"
18
19# WandB
20wandb_project: huggingface
21wandb_entity:
22wandb_name: TQ-2.5-14B-Sugarquill-1
23
24# Data
25#chat_template: chatml
26#train_on_inputs: false
27group_by_length: false
28datasets:
29 - path: allura-org/sugarquill-10k
30 type: completion
31
32## Evaluation
33val_set_size: 0.01
34evals_per_epoch: 4
35eval_table_size:
36eval_max_new_tokens: 128
37
38# Technical aspects
39sequence_len: 8192
40save_safetensors: true
41saves_per_epoch: 2
42logging_steps: 1
43special_tokens:
44
45# Quantization
46bf16: auto
47fp16:
48tf32: false
49## For LoRA
50load_in_8bit: false
51load_in_4bit: false
52
53# LoRA
54peft_use_rslora: true
55peft_use_dora: false # better but slower
56adapter: lora # lora or qlora
57lora_model_dir:
58lora_r: 64 # 64 is optimal for most trains on instruct
59lora_alpha: 32
60lora_dropout: 0.1
61lora_target_linear: true
62lora_fan_in_fan_out:
63lora_target_modules:
64# - embed_tokens
65# - lm_head
66
67#loraplus_lr_ratio: 8 # works to converge faster but is kinda cancer bc makes model unstable
68#loraplus_lr_embedding:
69
70# Training hyperparameters
71# max_steps:
72num_epochs: 2
73
74# Anti Overfit and Stability
75weight_decay: 0.01
76max_grad_norm: 1.0
77
78## Learning Rate
79warmup_ratio: 0.05
80learning_rate: 0.00003
81lr_scheduler: cosine
82#lr_scheduler_kwargs:
83# min_lr: 0.0000024
84optimizer: paged_ademamix_8bit # usually adamw_torch or paged_adamw_8bit
85
86## Batch Size
87gradient_accumulation_steps: 8 # More effective batch size - stabler train, usually. MBS also speeds it up.
88micro_batch_size: 1 # Batch size per gpu = micro_batch_size * gradient_accumulation_steps
89eval_batch_size: 1
90
91# Optimizations
92pad_to_sequence_len: true
93sample_packing: true
94eval_sample_packing: false
95flash_attention: true
96xformers_attention:
97gradient_checkpointing: "unsloth"
98gradient_checkpointing_kwargs:
99 use_reentrant: true
100local_rank:
101deepspeed: /home/kearm/axolotl/deepspeed_configs/zero3_bf16.json # Only use with multi gpu # _bf16_cpuoffload_all
102# fsdp:
103# - full_shard
104# - auto_wrap
105# fsdp_config:
106# fsdp_limit_all_gathers: true
107# fsdp_sync_module_states: true
108# fsdp_offload_params: true
109# fsdp_use_orig_params: false
110# fsdp_cpu_ram_efficient_loading: true
111# fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
112# fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
113# fsdp_state_dict_type: FULL_STATE_DICT
114# fsdp_sharding_strategy: FULL_SHARD
115# Misc
116early_stopping_patience:
117debug: