πŸ“‹ Model Description

license: mit base_model:
  • zerofata/MS3.2-PaintedFantasy-v4.1-24B





PaintedFantasy



















Painted Fantasy v4.1


Magistral Small 2509 24B





image






Overview






This is an uncensored model intended to excel at creative character driven RP / ERP.


Right after releasing v4 I noticed a bunch of repetition. Go figure. v4.1 is my first stab at trying to actively tailor the dataset towards weeding this out. Compared to v4, the only difference is heavy filtering and rewriting assistant messages identified as repetitive.


Repetition isn't fixed, but it's improved. The model still likes patterns, but at least seems capable of occasionally breaking these itself.














SillyTavern Settings





Recommended Roleplay Format





>
Actions:
In plaintext


>
Dialogue:
"In quotes"


>
Thoughts:
In asterisks





Recommended Samplers





>
Temp:
0.8


>
MinP:
0.05 - 0.075


>
TopP:
0.95 - 1.00





Instruct



Mistral v7 Tekken















Quantizations





GGUF





>
iMatrix





EXL3





>
3.0, 4.0, 5.0, 6.0bpw















Creation Process





Creation Process: SFT > DPO


SFT on approx 25 million tokens (17.5 million trainable). Datasets included SFW / NSFW RP, stories, NSFW reddit writing prompts, creative instruct & chat data.


90% of the dataset is without thinking, 10% included thinking, using the [THINK][/THINK] tags.


All RP data and synthetic stories went through rewriting with GLM 4.7 using hand edited examples as guidelines to improve the response. Rewritten responses were discarded if they failed to reduce the slop score for the message. This reduced the slop by about 25% for each RP / story dataset and made the model noticably more creative with some of its descriptions.


Assistant messages were checked for repetition in RP conversations via embeddings and word frequency checking across multi-turn conversations. Specific messages were rewritten and conversations that still showed high repetition were filtered.


DPO was expanded to include non creative datasets. My usual RP DPO dataset (also rewritten) was included along with cybersecurity and two partial subsets of general assistant / chat preference datasets to help stabalize the model. This worked pretty well. While creativity did take a small hit, enough remained that the improved logic resulted in a notably improved model (IMO).


Using embeddings, DPO samples where the chosen showed a higher similarity to the conversation than the rejected were removed, to ensure DPO doesn't encourage repetition.







>
Axolotl configs




Not optimized for cost / performance efficiency, YMMV.


SFT (4*H200)


base_model: Darkhn/Magistral-2509-24B-Text-Only



tokenizerusemistral_common: true



plugins:
  - axolotl.integrations.cutcrossentropy.CutCrossEntropyPlugin



loadin8bit: false
loadin4bit: false
deepspeed: deepspeed_configs/zero1.json



datasets:
  - path: ./data/nothink_dataset.jsonl
    type: chat_template
  - path: ./data/think_dataset.jsonl
    type: chat_template
datasetpreparedpath: lastrunprepared2
valsetsize: 0.01
output_dir: ./Magi-24B-SFT-v3-10



adapter: lora
peftuserslora: true
loramodeldir:



sequence_len: 10496
sample_packing: true
padtosequence_len: true



lora_r: 256
lora_alpha: 16
lora_dropout: 0.05
loratargetlinear: true



wandb_project: Magi-SFT-24B
wandb_name: Magi-24B-SFT-v3-10



gradientaccumulationsteps: 1
microbatchsize: 4
num_epochs: 2
optimizer: adamwbnb8bit
lr_scheduler: cosine



learning_rate: 1.5e-5
weight_decay: 0.01
maxgradnorm: 2.0



bf16: auto
tf32: false



gradient_checkpointing: true
resumefromcheckpoint:
logging_steps: 1
flash_attention: true



warmup_ratio: 0.05
evalsperepoch: 3
savesperepoch: 2



DPO (4*H200)


# ====================
# MODEL CONFIGURATION
# ====================
base_model: ApocalypseParty/Magi-24B-SFT-v3-10
model_type: MistralForCausalLM
tokenizer_type: AutoTokenizer
chattemplate: mistralv7_tekken



# ====================
# RL/DPO CONFIGURATION
# ====================
rl: dpo
rl_beta: 0.07



# ====================
# DATASET CONFIGURATION
# ====================
datasets:
  - path: ./data/dpoms32rewrittenhandcrafteddataset.jsonl
    type: chat_template.default
    field_messages: messages
    field_chosen: chosen
    field_rejected: rejected
    messagepropertymappings:
      role: role
      content: content
    roles:
      system: ["system"]
      user: ["user"]
      assistant: ["assistant"]
  - path: ./data/dpochubapprovedrewrittendataset_partial.jsonl
    type: chat_template.default
    field_messages: messages
    field_chosen: chosen
    field_rejected: rejected
    messagepropertymappings:
      role: role
      content: content
    roles:
      system: ["system"]
      user: ["user"]
      assistant: ["assistant"]
  - path: ./data/dposecureprogramming_dataset.jsonl
    type: chat_template.default
    field_messages: messages
    field_chosen: chosen
    field_rejected: rejected
    messagepropertymappings:
      role: role
      content: content
    roles:
      system: ["system"]
      user: ["user"]
      assistant: ["assistant"]
  - path: ./data/dpowildchatms32_chunk1.jsonl
    type: chat_template.default
    field_messages: messages
    field_chosen: chosen
    field_rejected: rejected
    messagepropertymappings:
      role: role
      content: content
    roles:
      system: ["system"]
      user: ["user"]
      assistant: ["assistant"]
  - path: ./data/dpoultrafeedbackchunk1.jsonl
    type: chat_template.default
    field_messages: messages
    field_chosen: chosen
    field_rejected: rejected
    messagepropertymappings:
      role: role
      content: content
    roles:
      system: ["system"]
      user: ["user"]
      assistant: ["assistant"]



datasetpreparedpath: ./dpo_data4
trainoninputs: false  # Only train on assistant responses



# ====================
# QLORA CONFIGURATION
# ====================
adapter: lora
loadin8bit: false
lora_r: 128
lora_alpha: 16
peftuserslora: true
lora_dropout: 0.1
loratargetlinear: true
# loramodulesto_save:  # Uncomment only if you added NEW tokens



# ====================
# TRAINING PARAMETERS
# ====================
num_epochs: 1
microbatchsize: 2
gradientaccumulationsteps: 4
learning_rate: 2e-6
optimizer: adamwtorchfused
lr_scheduler: cosine
warmup_ratio: 0.05
weight_decay: 0.01
maxgradnorm: 1.0



# ====================
# SEQUENCE CONFIGURATION
# ====================
sequence_len: 10756
padtosequence_len: true



# ====================
# HARDWARE OPTIMIZATIONS
# ====================
bf16: auto
tf32: false
flash_attention: true
gradient_checkpointing: offload



plugins:
  - axolotl.integrations.liger.LigerPlugin
  - axolotl.integrations.cutcrossentropy.CutCrossEntropyPlugin
cutcrossentropy: true
liger_rope: true
ligerrmsnorm: true
ligerlayernorm: true
ligergluactivation: true
ligercrossentropy: false  # Cut Cross Entropy overrides this
ligerfusedlinearcrossentropy: false  # Cut Cross Entropy overrides this
deepspeed: deepspeed_configs/zero1.json



# ====================
# CHECKPOINTING
# ====================
evalsperepoch: 1
savesperepoch: 6
loadbestmodelatend: true
metricforbestmodel: evalloss
greaterisbetter: false



# ====================
# LOGGING & OUTPUT
# ====================
output_dir: ./Magi-24B-SFT-v3-10-DPO-9
logging_steps: 1
save_safetensors: true



# ====================
# WANDB TRACKING
# ====================
wandb_project: Magi-24B-DPO
wandb_name: Magi-24B-SFT-v3-10-DPO-9












πŸ“‚ GGUF File List

πŸ“ Filename πŸ“¦ Size ⚑ Download
MS3.2-PaintedFantasy-v4.1-24B-IQ3_M.gguf
LFS Q3
9.92 GB Download
MS3.2-PaintedFantasy-v4.1-24B-IQ4_XS.gguf
LFS Q4
11.88 GB Download
MS3.2-PaintedFantasy-v4.1-24B-Q3_K_M.gguf
LFS Q3
10.69 GB Download
MS3.2-PaintedFantasy-v4.1-24B-Q4_K_L.gguf
LFS Q4
13.81 GB Download
MS3.2-PaintedFantasy-v4.1-24B-Q4_K_M.gguf
Recommended LFS Q4
13.35 GB Download
MS3.2-PaintedFantasy-v4.1-24B-Q5_K_M.gguf
LFS Q5
15.61 GB Download
MS3.2-PaintedFantasy-v4.1-24B-Q6_K.gguf
LFS Q6
18.02 GB Download
MS3.2-PaintedFantasy-v4.1-24B-Q8_0.gguf
LFS Q8
23.33 GB Download
MS3.2-PaintedFantasy-v4.1-24B-bf16.gguf
LFS FP16
43.92 GB Download