This is a storywriting and roleplay model with a significant amount of self generated long context multiturn roleplay.
I downloaded a bit under a thousand cards from chub.ai, and created a synthetic roleplay for each card. I batched as many turns as I could in 4k token chunks in order to maintain coherency over longer context. There was a lot of cleaning and validation between each batch, so a lot of examples were "lost," but the final output seems to be very good quality. The longest conversation is about 20k tokens, and I plan to extend this further as well as broaden the dataset with more examples. The first 4k tokens were generated with Command-R-Plus, with the remainder generated with byroneverson/Mistral-Small-Instruct-2409-abliterated.
Next, I downloaded the prompt backup from this site, and used them as a seed for some storywriting data:
https://aetherroom.club/whats-new#backup-update
I went over it twice with Command-R-Plus. The first time, having it basically write the first draft of the output, the second improving and extending the length of the original output.
Also included was a subset of the following datasets:
anthracite-org/stheno-filtered-v1.1
anthracite-org/kalo_misc_part2
anthracite-org/kalo_opus_misc_240827
anthracite-org/kalo-opus-instruct-22k-no-refusal
Chaser-cz/sonnet35-charcard-roleplay-sharegpt
(A very small subset) jondurbin/airoboros-3.2
And some various other data, viewable at openerotica/mixed-rp
Every line of data was run through a large model in order to filter for low quality, repetition, and underage content.
base_model: mistralai/Mistral-Nemo-Base-2407
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
load_in_8bit: false
load_in_4bit: true
strict: false
datasets:
- path: openerotica/mixed-rp
type: sharegpt
conversation: chatml
chat_template: chatml
adapter: qlora
lora_r: 128
lora_alpha: 256
lora_modules_to_save: [embed_tokens, lm_head]
lora_dropout: 0.05
lora_target_linear: true
lora_target_modules:
- gate_proj
- down_proj
- up_proj
- q_proj
- v_proj
- k_proj
- o_proj
dataset_prepared_path:
val_set_size: 0.01
output_dir: /workspace/axolotl/mixed-rp-mistral-nemo
sequence_len: 20000
sample_packing: true
pad_to_sequence_len: true
wandb_project: mistral-2
wandb_watch:
wandb_run_id:
wandb_log_model:
gradient_accumulation_steps: 2
micro_batch_size: 1
num_epochs: 1
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 1e-5
train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false
gradient_checkpointing: true
gradient_checkpointing_kwargs:
use_reentrant: false
early_stopping_patience:
resume_from_checkpoint:
logging_steps: 1
xformers_attention:
flash_attention: true
warmup_steps: 100
evals_per_epoch: 4
eval_table_size:
saves_per_epoch: 1
save_total_limit: 2
save_steps:
debug:
deepspeed:
weight_decay: 0.1
special_tokens:
eos_token: "<|im_end|>"
pad_token: "<pad>"
bos_token: "<s>"
unk_token: "<unk>"
tokens:
- "<|im_start|>"
# fsdp:
# - full_shard
# - auto_wrap
# fsdp_config:
# fsdp_limit_all_gathers: true
# fsdp_sync_module_states: true
# fsdp_offload_params: true
# fsdp_use_orig_params: false
# fsdp_cpu_ram_efficient_loading: true
# fsdp_transformer_layer_cls_to_wrap: MixtralSparseMoeBlock
# fsdp_state_dict_type: FULL_STATE_DICT
# fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
# fsdp_sharding_strategy: FULL_SHARD
# fsdp_forward_prefetch: false
# fsdp_backward_prefetch: BACKWARD_PRE