This is the eighth in 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 IntervitensInc/Llama-3.1-Minitron-4B-Width-Base-chatml.

Prompting

Model has been Instruct tuned with the ChatML formatting. A typical input would look like this:

"""<|im_start|>system
system prompt<|im_end|>
<|im_start|>user
Hi there!<|im_end|>
<|im_start|>assistant
Nice to meet you!<|im_end|>
<|im_start|>user
Can I ask a question?<|im_end|>
<|im_start|>assistant
"""

Support

Upstream support has been merged, so these quants work out of the box now!

old instructions before PR

To run inference on this model, you'll need to use Aphrodite, vLLM or EXL2/tabbyAPI, as llama.cpp hasn't yet merged the required pull request to fix the llama3.1 rope_freqs issue with custom head dimensions.

However, you can work around this by quantizing the model yourself to create a functional GGUF file. Note that until this PR is merged, the context will be limited to 8k tokens.

To create a working GGUF file, make the following adjustments:

Remove the "rope_scaling": {} entry from config.json
Change "max_position_embeddings" to 8192 in config.json

These modifications should allow you to use the model with llama.cpp, albeit with the mentioned context limitation.

axolotl config

See axolotl config

axolotl version: 0.4.1

yaml

base_model: IntervitensInc/Llama-3.1-Minitron-4B-Width-Base-chatml
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: anthracite-org/Gryphe-3.5-16k-Subset
    type: sharegpt
    conversation: chatml
  - path: Epiculous/Synthstruct-Gens-v1-Filtered-n-Cleaned
    type: sharegpt
    conversation: chatml
  - path: anthracite-org/Stheno-Data-Filtered
    type: sharegpt
    conversation: chatml
  - path: Epiculous/SynthRP-Gens-v1-Filtered-n-Cleaned
    type: sharegpt
    conversation: chatml
  - path: lodrick-the-lafted/NopmWritingStruct
    type: sharegpt
    conversation: chatml
  - path: anthracite-org/kalo-opus-instruct-22k-no-refusal
    type: sharegpt
    conversation: chatml

chat_template: chatml

val_set_size: 0.01
output_dir: ./outputs/out

adapter:
lora_r:
lora_alpha:
lora_dropout:
lora_target_linear:

sequence_len: 16384
# sequence_len: 32768
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 32
micro_batch_size: 1
num_epochs: 2
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.00002
weight_decay: 0.05

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

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

warmup_ratio: 0.1
evals_per_epoch: 4
eval_table_size:
eval_max_new_tokens: 128
saves_per_epoch: 1

debug:
deepspeed:
fsdp:
fsdp_config:

special_tokens:
  pad_token: <|finetune_right_pad_id|>

Credits

This model has been a team effort, and the credits goes to all members of Anthracite.

Training

The training was done for 2 epochs. We used 2 x RTX 6000s GPUs graciously provided by Kubernetes_Bad for the full-parameter fine-tuning of the model.

Safety

...

anthracite-org/magnum-v2-4b-gguf

This repo contains GGUF quants of the model. If you need the original weights, please find them here.

The quants were made with the mentioned PR merged.