metadata

library_name: transformers
license: apache-2.0
base_model:
  - THUDM/GLM-4-32B-0414
tags:
  - roleplay
  - conversational
  - axolotl

Remnant GLM4 32b (series 1)

English | 简体中文

There's a wisp of dust in the air. It feels like its from a bygone era, but you don't know where from. It lands on your tongue. It tastes nice.

Remnant is a series of finetuned LLMs focused on SFW and NSFW roleplaying and conversation.

Quants

GGUF (IF YOU ARE USING KOBOLD.CPP, PLEASE USE --nobostoken.):

bartowski's imatrix quants

EXL3:

Todo!

EXL2 (if it ever gets supported):

Todo!

MISC:

Todo!

Recommended Settings

Chat template: GLM4
Samplers:

1.0 temp
0.1 min_p

An example Sillytavern template is available here.

Credits

Humongous thanks to Allura, ilya <3
Big thanks to the developers of Axolotl (whose training framework I used), Zhipu (whose model I used), Prime Intellect (whose GPUs I used), and my bank (whose debit card I used)

Misc

See axolotl config

axolotl version: 0.8.0

# === Model Configuration ===
base_model: THUDM/GLM-4-32B-0414  # e.g. "mistralai/Mistral-Small-24B-Instruct-2501"
load_in_8bit: false
load_in_4bit: true

# === Training Setup ===
num_epochs: 2
micro_batch_size: 3
gradient_accumulation_steps: 2
sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true

# === Hyperparameter Configuration ===
optimizer: adamw_8bit
# Apollo-mini configuration:
#optim_args: "proj=random,rank=1,scale=128.0,scale_type=tensor,update_proj_gap=200"
# Regular Apollo configuration:
# optim_args: 
#optim_target_modules: all_linear
learning_rate: 1e-5
lr_scheduler: rex
weight_decay: 0.01
warmup_ratio: 0.05

# === LoRA Configuration ===
adapter: qlora
lora_r: 16
lora_alpha: 32
lora_dropout: 0.25
lora_target_modules:
lora_target_linear: true

# === Data Configuration ===
datasets:
  - path: allura-org/inkmix-v3.0
    type: chat_template
    split: train
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    train_on_eos: all

dataset_prepared_path: last_run_prepared
chat_template: jinja
chat_template_jinja: |
  [gMASK]<sop>{%- for msg in messages %}{%- if msg.role == 'system' %}<|system|>
  {{ msg.content }}{%- elif msg.role == 'user' %}<|user|>
  {{ msg.content }}{%- elif msg.role == 'assistant' %}<|assistant|>
  {{ msg.content }}{%- endif %}{%- endfor %}{% if add_generation_prompt %}<|assistant|>{% else %}<|user|>{% endif %}

# === Plugins ===
plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

# === Hardware Optimization ===
gradient_checkpointing: offload
gradient_checkpointing_kwargs:
  use_reentrant: false
cut_cross_entropy: true
deepspeed: deepspeed_configs/zero3_bf16.json

# === Wandb Tracking ===
wandb_project: glm4-32b-inkmix-v3

# === Checkpointing ===
saves_per_epoch: 2
save_total_limit: 3

# === Advanced Settings ===
output_dir: /workspace/ckpts
bf16: auto
flash_attention: true
train_on_inputs: false
group_by_length: false
logging_steps: 1
trust_remote_code: true