Delta-Vector/Odin-9B

A earlier checkpoint of Magnum V4 9B model, using the same configuration as Tor-8B / Darkens-8B but on Gemma rather then Nemo-8B, A finetune made for creative writing and roleplay tasks, Finetuned ontop of the base Gemma2 9B model, I trained the model for 4 epochs, with the 4 epoch checkpoint becoming the Magnum model and the 2 epoch checkpoint becoming my own personal release. This model aims to have good prose and writing while not as Suggestive as Magnum models usually are, along with keeping some of the intelligence that was nice to have with the Gemma2 family.

Quants

GGUF: https://huggingface.co/Delta-Vector/Odin-9B-GGUF

EXL2: https://huggingface.co/Delta-Vector/Odin-9B-EXL2

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
"""

System Prompting

I would highly recommend using Sao10k's Euryale System prompt, But the "Roleplay Simple" system prompt provided within SillyTavern will work aswell. Also Use 0.02 minp for the models, The model may act dumb or otherwise stupid without it.

Currently, your role is {{char}}, described in detail below. As {{char}}, continue the narrative exchange with {{user}}.

<Guidelines>
• Maintain the character persona but allow it to evolve with the story.
• Be creative and proactive. Drive the story forward, introducing plotlines and events when relevant.
• All types of outputs are encouraged; respond accordingly to the narrative.
• Include dialogues, actions, and thoughts in each response.
• Utilize all five senses to describe scenarios within {{char}}'s dialogue.
• Use emotional symbols such as "!" and "~" in appropriate contexts.
• Incorporate onomatopoeia when suitable.
• Allow time for {{user}} to respond with their own input, respecting their agency.
• Act as secondary characters and NPCs as needed, and remove them when appropriate.
• When prompted for an Out of Character [OOC:] reply, answer neutrally and in plaintext, not as {{char}}.
</Guidelines>

<Forbidden>
• Using excessive literary embellishments and purple prose unless dictated by {{char}}'s persona.
• Writing for, speaking, thinking, acting, or replying as {{user}} in your response.
• Repetitive and monotonous outputs.
• Positivity bias in your replies.
• Being overly extreme or NSFW when the narrative context is inappropriate.
</Forbidden>

Follow the instructions in <Guidelines></Guidelines>, avoiding the items listed in <Forbidden></Forbidden>.

Axolotl config

See axolotl config

Axolotl version: 0.4.1

base_model: /workspace/data/gemma-2-9b-chatml
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

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

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: anthracite-org/c2_logs_16k_llama_v1.1
    type: sharegpt
    conversation: chatml
  - path: NewEden/Claude-Instruct-5K
    type: sharegpt
    conversation: chatml  
  - path: anthracite-org/kalo-opus-instruct-22k-no-refusal
    type: sharegpt
    conversation: chatml
  - path: Epiculous/SynthRP-Gens-v1.1-Filtered-n-Cleaned
    type: sharegpt
    conversation: chatml
  - path: lodrick-the-lafted/kalo-opus-instruct-3k-filtered
    type: sharegpt
    conversation: chatml
  - path: anthracite-org/nopm_claude_writing_fixed
    type: sharegpt
    conversation: chatml
  - path: Epiculous/Synthstruct-Gens-v1.1-Filtered-n-Cleaned
    type: sharegpt
    conversation: chatml
  - path: anthracite-org/kalo_opus_misc_240827
    type: sharegpt
    conversation: chatml
  - path: anthracite-org/kalo_misc_part2
    type: sharegpt
    conversation: chatml
chat_template: chatml
shuffle_merged_datasets: false
default_system_message: "You are a helpful assistant that responds to the user."
dataset_prepared_path: /workspace/data/9b-fft-data
val_set_size: 0.0
output_dir: /workspace/data/9b-fft-out

sequence_len: 8192
sample_packing: true
eval_sample_packing: false
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: 9b-Nemo-config-fft
wandb_entity:
wandb_watch:
wandb_name: attempt-01
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 1
num_epochs: 4
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 0.00001

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

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

warmup_steps: 10
evals_per_epoch:
eval_table_size:
eval_max_new_tokens:
saves_per_epoch: 1
debug:
deepspeed: deepspeed_configs/zero3_bf16.json
weight_decay: 0.001
fsdp:
fsdp_config:
special_tokens:
  pad_token: <pad>

Credits

Thank you to Lucy Knada, Kalomaze, Kubernetes Bad and the rest of Anthracite (But not Alpin.)

Training

The training was done for 4 epochs. We used 8 x H100s GPUs graciously provided by Lucy Knada for the full-parameter fine-tuning of the model.

Safety

Nein.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	24.65
IFEval (0-Shot)	36.92
BBH (3-Shot)	34.83
MATH Lvl 5 (4-Shot)	12.54
GPQA (0-shot)	12.19
MuSR (0-shot)	17.56
MMLU-PRO (5-shot)	33.85