See axolotl config

axolotl version: 0.13.0.dev0

yaml

1# 1. Base Model & Tokenizer 2base_model: mistralai/Mistral-7B-Instruct-v0.3 3model_type: MistralForCausalLM 4tokenizer_type: LlamaTokenizer 5hub_model_id: AiAF/Adapter_Mistral-7B-Instruct-v0.3-co-sft-qlora 6 7# 2. LoRA / QLoRA Configuration 8load_in_8bit: false 9load_in_4bit: true 10adapter: qlora 11# Matching gemma2 config for potentially better results 12lora_r: 64 13lora_alpha: 128 14lora_dropout: 0.05 15lora_target_linear: true 16lora_target_modules: 17 - gate_proj 18 - down_proj 19 - up_proj 20 - q_proj 21 - v_proj 22 - k_proj 23 - o_proj 24 25# 3. Dataset Configuration 26datasets: 27 - path: . 28 type: chat_template 29 data_files: ./co-sft-dataset.jsonl 30 field_messages: conversations 31 message_property_mappings: 32 role: from 33 content: value 34 chat_template: jinja 35 chat_template_jinja: | 36 {{ bos_token }} 37 38 {# 1) Collect all system messages and join them #} 39 {% set sys = ( 40 messages 41 | selectattr('role', 'equalto', 'system') 42 | map(attribute='content') 43 | list 44 ) %} 45 {% set sys_joined = sys | reject('equalto', None) | map('trim') | select | join('\n\n') %} 46 47 {# 2) Emit conversation without 'system' role; inject system text into the FIRST human turn #} 48 {% set injected = namespace(done=false) %} 49 {% for m in messages if m['role'] != 'system' %} 50 {% if m['role'] == 'human' %} 51 {% if not injected.done and sys_joined %} 52 {{ '[INST] ' + (sys_joined ~ '\n\n' ~ (m['content'] | trim)) + ' [/INST]' }} 53 {% set injected.done = true %} 54 {% else %} 55 {{ '[INST] ' + (m['content'] | trim) + ' [/INST]' }} 56 {% endif %} 57 {% elif m['role'] == 'assistant' %} 58 {{ ' ' + (m['content'] | trim) + eos_token }} 59 {% endif %} 60 {% endfor %} 61 62 roles_to_train: ["assistant"] 63 64# 4. Training Parameters 65sequence_len: 2048 66sample_packing: true 67val_set_size: 0.05 68num_epochs: 10 69dataset_prepared_path: last_run_prepared 70gradient_accumulation_steps: 1 #4 71micro_batch_size: 32 #2 72optimizer: adamw_bnb_8bit 73lr_scheduler: cosine 74learning_rate: 0.0002 75bf16: true 76tf32: true 77gradient_checkpointing: true 78 79# 5. Saving and Evaluation Strategy 80output_dir: ./outputs/sft/Mistral-7B-Instruct-v0.3-co-sft-qlora 81logging_steps: 10 82evals_per_epoch: 5 83saves_per_epoch: 10 84save_total_limit: 100 85 86# 6. W&B Logging 87wandb_project: "co-sft" 88wandb_name: "Mistral-7B-Instruct-v0.3-co-sft-qlora-co-sft" 89wandb_log_model: "false" 90wandb_run_id: "Mistral-7B-Instruct-v0.3-co-sft-qlora" 91 92# 7. Special Tokens 93special_tokens: 94 bos_token: "<s>" 95 eos_token: "</s>" 96 unk_token: "<unk>" 97

Adapter_Mistral-7B-Instruct-v0.3-co-sft-qlora

• Loss: 2.9397
• Memory/max Active (gib): 46.02
• Memory/max Allocated (gib): 46.02
• Memory/device Reserved (gib): 53.04

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

• learning_rate: 0.0002
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_BNB with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 41
• training_steps: 1390

Training results

Framework versions

• PEFT 0.17.1
• Transformers 4.56.1
• Pytorch 2.7.1+cu126
• Datasets 4.0.0
• Tokenizers 0.22.1