---
library_name: peft
license: apache-2.0
base_model: mistralai/Mistral-Small-3.2-24B-Instruct-2506
tags:
- axolotl
- generated_from_trainer
---

# MS3.2-24B-Magnum-Diamond-LoRA

Magnum "Diamond" in reference to the intense heat and pressure (generated through matrix multiplications) needed to turn the coal-esque material of dry, assistant-tuned models into creative writing gems!

This model is finetuned from a text-only conversion of [mistralai/Mistral-Small-3.2-24B-Instruct-2506](https://huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506) as an rsLoRA adapter. It uses the same data mix as [Doctor-Shotgun/L3.3-70B-Magnum-v5-SFT-Alpha](https://huggingface.co/Doctor-Shotgun/L3.3-70B-Magnum-v5-SFT-Alpha), however with pre-tokenization and modifications to the custom loss masking.
The goal was to re-create the model at a smaller, more consumer-friendly size.
This model should perform competently with or without prepending character names, and with or without prefill.

The objective, as with the other Magnum models, is to emulate the prose style and quality of the Claude 3 Sonnet/Opus series of models on a local scale, so don't be surprised to see "Claude-isms" in its output.

This is a minor version update over [Doctor-Shotgun/MS3.1-24B-Magnum-Diamond-LoRA](https://huggingface.co/Doctor-Shotgun/MS3.1-24B-Magnum-Diamond-LoRA) utilizing the new official instruct model from June 2025.

[Merged full model](https://huggingface.co/Doctor-Shotgun/MS3.2-24B-Magnum-Diamond)

## Intended uses and limitations

This model is intended for creative writing and roleplay purposes.
It may show biases similar to those observed in contemporary LLM-based roleplay, in addition to those exhibited by the Claude 3 series of models and the base model.
All outputs should be considered fiction, as this model is not intended to provide factual information or advice. 

## Training procedure

[WandB](https://wandb.ai/gum1h0x/24b-magnum-lora/runs/3zudxeg3?nw=nwuseradrianjuliusbeck)

There was a weird loss spike of unclear significance on one sample that was not seen using the same dataset on Mistral Small 3.1 Instruct, but the resulting model appears to be sane.

[<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/axolotl-ai-cloud/axolotl)
<details><summary>See axolotl config</summary>

axolotl version: `0.9.2`
```yaml
base_model: anthracite-core/Mistral-Small-3.2-24B-Instruct-2506-Text-Only
#base_model_ignore_patterns: "consolidated.safetensors"
# optionally might have model_type or tokenizer_type
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
# Automatically upload checkpoint and final model to HF
hub_model_id: NewEden/magnum-v5-sft-prototype-ms3.2-lora
hub_strategy: "all_checkpoints"
push_dataset_to_hub:
hf_use_auth_token: true

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: NewEden/magnum-v5-sft-proto-mistral-v7-tekken-rev1-32k
    ds_type: parquet
    type:
shuffle_merged_datasets: true
dataset_prepared_path: ./magnum-24b-data
val_set_size: 0.0
output_dir: ./magnum-24b-lora-out

plugins:
  - axolotl.integrations.liger.LigerPlugin
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
liger_rope: true
liger_rms_norm: true
liger_layer_norm: true
liger_glu_activation: true
liger_fused_linear_cross_entropy: false
cut_cross_entropy: true

sequence_len: 32768
sample_packing: true
eval_sample_packing: true
pad_to_sequence_len: true

adapter: lora
lora_model_dir:
lora_r: 128
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
peft_use_rslora: true
lora_modules_to_save:
  - embed_tokens
  - lm_head

wandb_project: 24b-magnum-lora
wandb_entity:
wandb_watch:
wandb_name: 24b-magnum-lora-mistral-3.2
wandb_log_model:

gradient_accumulation_steps: 16
micro_batch_size: 1
num_epochs: 2
optimizer: paged_ademamix_8bit
lr_scheduler: cosine
learning_rate: 2e-5
max_grad_norm: 1.0

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
s2_attention:

warmup_steps: 40
evals_per_epoch:
eval_table_size:
eval_max_new_tokens:
saves_per_epoch: 2
debug:
deepspeed:
weight_decay: 0.01
fsdp:
fsdp_config:
special_tokens:

```

</details><br>

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 1
- eval_batch_size: 1
- seed: 42
- gradient_accumulation_steps: 16
- total_train_batch_size: 16
- optimizer: Use paged_ademamix_8bit and the args are:
No additional optimizer arguments
- lr_scheduler_type: cosine
- lr_scheduler_warmup_steps: 40
- num_epochs: 2.0

### Framework versions

- PEFT 0.15.2
- Transformers 4.51.3
- Pytorch 2.7.1+cu128
- Datasets 3.5.1
- Tokenizers 0.21.1