anthracite-org/magnum-v4-123b

This is a series of models designed to replicate the prose quality of the Claude 3 models, specifically Sonnet and Opus.

This model is fine-tuned on top of mistralai/Mistral-Large-Instruct-2407.

Prompting

A typical input would look like this:

<s>[INST] SYSTEM MESSAGE\nUSER MESSAGE[/INST] ASSISTANT MESSAGE</s>[INST] USER MESSAGE[/INST]

SillyTavern templates

Below are Instruct and Context templates for use within SillyTavern.

context template

default SillyTavern template works fine

instruct template

default SillyTavern template works fine

Axolotl config

See axolotl config

base_model: mistralai/Mistral-Large-Instruct-2407
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

plugins:
  - axolotl.integrations.liger.LigerPlugin
liger_rope: true
liger_rms_norm: true
liger_swiglu: true
liger_fused_linear_cross_entropy: true

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: anthracite-org/c2_logs_16k_mistral-large_v1.2
    type: sharegpt
    conversation: mistral
  - path: anthracite-org/kalo-opus-instruct-22k-no-refusal
    type: sharegpt
    conversation: mistral
  - path: lodrick-the-lafted/kalo-opus-instruct-3k-filtered
    type: sharegpt
    conversation: mistral
  - path: anthracite-org/nopm_claude_writing_fixed
    type: sharegpt
    conversation: mistral
  - path: anthracite-org/kalo_opus_misc_240827
    type: sharegpt
    conversation: mistral
  - path: anthracite-org/kalo_misc_part2
    type: sharegpt
    conversation: mistral
#chat_template: chatml
shuffle_merged_datasets: true
#default_system_message: "You are an assistant that responds to the user."
dataset_prepared_path: ./data/magnum-123b-data
val_set_size: 0.0
output_dir: ./data/123b-fft-out

sequence_len: 16384
sample_packing: true
pad_to_sequence_len: true

adapter:
lora_model_dir:
lora_r:
lora_alpha:
lora_dropout:
lora_target_linear:
lora_fan_in_fan_out:

wandb_project: 123b-magnum-fft
wandb_entity:
wandb_watch:
wandb_name: alter-attempt-04
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 1
num_epochs: 2
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0000015

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false

gradient_checkpointing: unsloth
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 40
evals_per_epoch:
eval_table_size:
eval_max_new_tokens:
saves_per_epoch: 2
debug:
deepspeed: deepspeed_configs/zero3_bf16.json
weight_decay: 0.01
fsdp:
fsdp_config:
special_tokens:

Credits

We'd like to thank Eric Hartford for sponsoring the compute for this train. We would also like to thank all members of Anthracite who made this finetune possible.

Datasets

• anthracite-org/c2_logs_16k_mistral-large_v1.2
• anthracite-org/kalo-opus-instruct-22k-no-refusal
• lodrick-the-lafted/kalo-opus-instruct-3k-filtered
• anthracite-org/nopm_claude_writing_fixed
• anthracite-org/kalo_opus_misc_240827
• anthracite-org/kalo_misc_part2

Training

We used 8x mi300x GPUs graciously provided by Eric Hartford for the full-parameter fine-tuning of the model.

Safety

...