Stardom

Mero Mero v2

Gemma4 31B 01 Overview

A finetune of Gemma 4 31B designed for creative tasks, particularly narrative RP. Intended to be a more creative version of G4-MeroMero-31B.

This model is the result of a lot of experimentation and learning. Trying to make Gemma 4 more creative without destroying the intelligence is... difficult. To put it mildly.

Heavily inspired by a few research papers, StoryScope: Investigating idiosyncrasies in AI fiction and particularly Elias in the Lighthouse, Again?. Measuring these narrative tics and attractors against simple prompts seems to be a good way to target the model's slop and kick start giving Gemma 4 some diversity: anything that repeatedly occurs across generations of such a generic prompt is something the model is overusing.

Compared to the original, swipes are notably more diverse and feel less like Gemma. RP slop is measurably lower (at least for the type of slop I measure). IFEval / GSM8K / MMLU-Pro are the same as stock with no obvious degradation. The only intelligence drop I've really noticed so far is when you get a swipe that goes a bit hot.

Supports both thinking and non thinking. Reasoning averages longer than stock Gemma 4, but shorter than MeroMero v1.

02 SillyTavern Settings Suggested Roleplay Format ActionsIn plaintext Dialogue"In quotes" Thoughts*In asterisks* Recommended Samplers Temp0.8 - 1.0 MinP0.05 Instruct Gemma 4 - Think Gemma 4 - NoThink 03 Quantizations GGUF iMatrix 04 Evaluation 05 Creation Process

Creation Process: SFT > Merge > GRPO > GRPO > on-policy SFT

Stage 1 — Diversity SFT. Stock Gemma 4 collapses hard on underspecified creative prompts ("Write a story." basically always gives clockmaker or memory related stories in a shop with Elias). Trained on ~4,000 short stories curated against the storyscope narrative prompts and found attractors. The dataset is a mix of human stories and synthetic stories from a set of frontier models, with diverse generation prompts swapped out for generic ones and filtered for quality. I also included some of the usual creative instruct and roleplay data. The model came out alright. Creative, but notably worse at instruction following with degraded intelligence. SLERP-merged back into the original instruct at t=0.5, which basically reverted it to stock Gemma 4 with slightly improved prose and creativity (similar to MeroMero).

Stage 2 — Creative GRPO (with think disabled). TRL GRPO (via Axolotl), 8 rollouts per prompt on the same bare prompts. Reward stack: LLM-judge diversity and coherence rewards, an attractor-marker penalty seeded from stock and then updated with whatever started appearing as new attractors during training, narrative-rate penalties and deterministic degeneracy guards (checking for non-Latin characters, joined words etc). 300 steps.

Stage 3 — RP logic GRPO (with think enabled). 100 further steps on multi-turn roleplay contexts: a thinking check to ensure it always parsed correctly, a logic-defect judge (DeepSeek-V4 Flash with a rubric), per-context attractor lists mined from k=8 baselines of the stage 2 model, and the same degeneracy checks as stage 2.

Stage 4 — On-policy multi-outcome RP SFT. ~3,300 samples the model wrote itself on roleplay contexts, filtered to keep only varied continuations. The model wasn't able to produce diverse results purely on-policy, so a natural response was generated first, then each sample went through a GLM-5.2 critique pipeline: error detection, plus steering ideas injected as OOC commands for re-generation (DeepSeek-V4-Pro / GLM-5.2 alternating). Everything was then filtered for errors, slop and any degeneracy as usual. Roughly 60% of samples include thinking. Last turn only.

Trained using Axolotl.

Stage 1 — Diversity SFT Config (Axolotl)

base_model: google/gemma-4-31B-it
 
plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
  - axolotl.integrations.liger.LigerPlugin
liger_layer_norm: true
liger_rope: true
liger_rms_norm: true
liger_glu_activation: true
liger_rms_norm_gated: true
strict: false
cut_cross_entropy: true
 
datasets:
  - path: ./data/diversity_sft_masked.jsonl
val_set_size: 0
output_dir: ./G4-31B-SFT-v10-2
 
sequence_len: 8192
pad_to_sequence_len: true
sample_packing: true
 
adapter: lora
lora_r: 64
lora_alpha: 64
peft_use_rslora: true
lora_dropout: 0.0
freeze_mm_modules: true
lora_target_modules: 'model.language_model.layers.[\d]+.(_checkpoint_wrapped_module.)?(mlp|self_attn).(up|down|gate|q|k|v|o)_proj'
 
gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 2
optimizer: adamw_torch_fused
lr_scheduler: cosine
learning_rate: 1e-5
max_grad_norm: 1.0
warmup_ratio: 0.1
weight_decay: 0.05
saves_per_epoch: 2
 
bf16: auto
tf32: true
 
# FA2 not supported
sdp_attention: true
flash_attention: false
 
fsdp_config:
  fsdp_version: 2
  offload_params: false
  cpu_ram_efficient_loading: false
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: Gemma4TextDecoderLayer
  state_dict_type: FULL_STATE_DICT
  sharding_strategy: FULL_SHARD
  reshard_after_forward: true
  activation_checkpointing: true

Stage 1 — Mergekit Config

merge_method: slerp
base_model: google/gemma-4-31B-it
models:
  - model: google/gemma-4-31B-it
  - model: ApocalypseParty/G4-31B-SFT-v10-2
parameters:
  t: 0.5
dtype: bfloat16

Stage 2 — Creative GRPO Config (Axolotl)

base_model: /workspace/models/configCA   # stage 1 output
 
rl: grpo
 
trl:
  reward_funcs:
    - rewards_g4.reward_judge_diversity
    - rewards_g4.reward_judge_coherence
    - rewards_g4.reward_attractor
    - rewards_g4.reward_narrative
    - rewards_g4.reward_sane
  reward_weights: [3.0, 3.0, 0.75, 1.0, 1.0]
  beta: 0.02
  num_generations: 8
  max_completion_length: 1600
  temperature: 1.0
  use_vllm: true
  scale_rewards: true
  loss_type: grpo
  epsilon: 0.2
  generation_kwargs:
    stop_token_ids: [1, 106, 50]
    top_k: 64
    top_p: 0.95
 
datasets:
  - path: /workspace/data/sft_train_final.jsonl
    type: ebft_chat.transform
 
sequence_len: 2048
micro_batch_size: 2
gradient_accumulation_steps: 4
max_steps: 200
 
learning_rate: 4.0e-6
optimizer: adamw_torch_fused
lr_scheduler: cosine
warmup_steps: 10
weight_decay: 0.01
 
adapter: lora
lora_r: 64
lora_alpha: 64
peft_use_rslora: true
lora_dropout: 0.0
freeze_mm_modules: true
lora_target_modules: 'model.language_model.layers.[\d]+.(_checkpoint_wrapped_module.)?(mlp|self_attn).(up|down|gate|q|k|v|o)_proj'
 
max_grad_norm: 1.0
bf16: auto
tf32: true
sdp_attention: true
flash_attention: false
gradient_checkpointing: true
 
# shipped weights use checkpoint-100 of this run

Stage 3 — RP Logic GRPO Config (Axolotl)

base_model: /workspace/models/r4b100   # stage 2 output
 
rl: grpo
 
trl:
  reward_funcs:
    - rewards_rp.reward_thinking   # format gate on the think block
    - rewards_rp.reward_logic      # constraint-grounded defect judge
    - rewards_rp.reward_attractor  # frozen per-context lists from stock k=8
    - rewards_rp.reward_sane       # deterministic glitch guards
  reward_weights: [2.0, 3.0, 1.0, 1.0]
  beta: 0.02
  num_generations: 8
  max_completion_length: 2560
  temperature: 1.0
  use_vllm: true
  scale_rewards: true
  loss_type: grpo
  epsilon: 0.2
  generation_kwargs:
    stop_token_ids: [1, 106, 50]
    top_k: 64
    top_p: 0.95
 
datasets:
  - path: /workspace/rp/rp3_train.jsonl
    type: ebft_chat.transform
 
sequence_len: 8192
micro_batch_size: 1
gradient_accumulation_steps: 8
max_steps: 100
 
learning_rate: 3.0e-6
optimizer: adamw_torch_fused
lr_scheduler: cosine
warmup_steps: 10
weight_decay: 0.01
 
adapter: lora
lora_r: 64
lora_alpha: 64
peft_use_rslora: true
lora_dropout: 0.0
freeze_mm_modules: true
lora_target_modules: 'model.language_model.layers.[\d]+.(_checkpoint_wrapped_module.)?(mlp|self_attn).(up|down|gate|q|k|v|o)_proj'
 
max_grad_norm: 1.0
bf16: auto
tf32: true
sdp_attention: true
flash_attention: false
gradient_checkpointing: true

Stage 4 — On-Policy Multi-Outcome SFT Config (Axolotl)

base_model: ApocalypseParty/G4-31B-r4b100-GRPO-rp100   # stage 3 output
 
plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
  - axolotl.integrations.liger.LigerPlugin
liger_layer_norm: true
liger_rope: true
liger_rms_norm: true
liger_glu_activation: true
liger_rms_norm_gated: true
strict: false
cut_cross_entropy: true
 
datasets:
  - path: ./data/g4_onpolicy_rp_masked.jsonl
val_set_size: 0
output_dir: ./G4-31B-r4b100-GRPO-rp100-sft
 
sequence_len: 8192
pad_to_sequence_len: true
sample_packing: true
 
adapter: lora
lora_r: 64
lora_alpha: 64
peft_use_rslora: false
lora_dropout: 0.0
freeze_mm_modules: true
lora_target_modules: 'model.language_model.layers.[\d]+.(_checkpoint_wrapped_module.)?(mlp|self_attn).(up|down|gate|q|k|v|o)_proj'
 
gradient_accumulation_steps: 2
micro_batch_size: 1
num_epochs: 1
optimizer: adamw_torch_fused
lr_scheduler: cosine
learning_rate: 4e-5
max_grad_norm: 1.0
warmup_ratio: 0.1
weight_decay: 0.05
saves_per_epoch: 2
 
bf16: auto
tf32: true
 
# FA2 not supported
sdp_attention: true
flash_attention: false
 
fsdp_config:
  fsdp_version: 2
  offload_params: false
  cpu_ram_efficient_loading: false
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: Gemma4TextDecoderLayer
  state_dict_type: FULL_STATE_DICT
  sharding_strategy: FULL_SHARD
  reshard_after_forward: true
  activation_checkpointing: true