danube2-upscale-1.1

Aim was to repair damage caused by duplicating the upscale with some additional training on completion from Cosmopedia.

Seemed to be converged at 50% epoch so I cut it off and used that adapter, which I hope actually did something because it wasn't a checkpoint.

eq_bench testing, as a quick reference, strongly suggests it did; but I'm not sure how much that one's just random on a small model like this.

It also seems to be generating completions much more smoothly than its predecessor, though, rather than getting stuck in a repeated word, which is certainly a good sign.

Nous evals:

Original model:

Axolotl config was something like this:

base_model: Lambent/danube2-upscale-1
model_type: MistralForCausalLM
tokenizer_type: LlamaTokenizer
trust_remote_code: false

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: HuggingFaceTB/cosmopedia-100k
    type: completion
dataset_prepared_path: prepared-pedia
val_set_size: 0.01
output_dir: ./qlora-out

sequence_len: 8192
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true

adapter: qlora
lora_model_dir:
lora_r: 128
lora_alpha: 128
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:

wandb_project: qlora-danube-upscale
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 1
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 0.0002

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

loss_watchdog_threshold: 5.0
loss_watchdog_patience: 3

warmup_steps: 10
evals_per_epoch: 4
saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 0.002
fsdp:
fsdp_config:
special_tokens: