anthracite-org/magnum-v4-22b

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 Mistral-Small-Instruct-2409.

Prompting

A typical input would look like this:

<s>[INST] SYSTEM MESSAGE
USER 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: /workspace/models/Mistral-Small-Instruct-2409
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

hub_model_id: anthracite-org/magnum-v4-22b-r4
hub_strategy: "all_checkpoints"
push_dataset_to_hub:
hf_use_auth_token: true

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

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: anthracite-org/c2_logs_32k_mistral-v3_v1.2_no_system
    type: custommistralv2v3
  - path: anthracite-org/kalo-opus-instruct-22k-no-refusal-no-system
    type: custommistralv2v3
  - path: anthracite-org/kalo-opus-instruct-3k-filtered-no-system
    type: custommistralv2v3
  - path: anthracite-org/nopm_claude_writing_fixed
    type: custommistralv2v3
  - path: anthracite-org/kalo_opus_misc_240827_no_system
    type: custommistralv2v3
  - path: anthracite-org/kalo_misc_part2_no_system
    type: custommistralv2v3
#chat_template: mistral_v2v3
shuffle_merged_datasets: true
#default_system_message: "You are an assistant that responds to the user."
dataset_prepared_path: /workspace/data/magnum-22b-data
val_set_size: 0.0
output_dir: /workspace/data/22b-r4-fft-out

sequence_len: 32768
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: 22b-magnum-fft
wandb_entity:
wandb_watch:
wandb_name: v4-r4-attempt-01
wandb_log_model:

gradient_accumulation_steps: 2
micro_batch_size: 1
num_epochs: 2
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.000004

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

gradient_checkpointing: true
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: 0.1
fsdp:
fsdp_config:
special_tokens:

Credits

We'd like to thank Recursal / Featherless for sponsoring the compute for this train, Featherless has been hosting our Magnum models since the first 72 B and has given thousands of people access to our models and helped us grow.

We would also like to thank all members of Anthracite who made this finetune possible.

Datasets

• anthracite-org/c2_logs_32k_mistral-v3_v1.2_no_system
• anthracite-org/kalo-opus-instruct-22k-no-refusal-no-system
• anthracite-org/kalo-opus-instruct-3k-filtered-no-system
• anthracite-org/nopm_claude_writing_fixed
• anthracite-org/kalo_opus_misc_240827_no_system
• anthracite-org/kalo_misc_part2_no_system

Training

The training was done for 2 epochs. We used 8xH100s GPUs graciously provided by Recursal AI / Featherless AI for the full-parameter fine-tuning of the model.

Safety

...

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	27.45
IFEval (0-Shot)	56.29
BBH (3-Shot)	35.55
MATH Lvl 5 (4-Shot)	17.60
GPQA (0-shot)	10.40
MuSR (0-shot)	13.43
MMLU-PRO (5-shot)	31.44