# ── Continued Pretraining: 7B on 8×A40 (48GB) ──

base_model: allenai/Olmo-3-1025-7B
tokenizer_type: AutoTokenizer

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

# ── Data ──
datasets:
  - path: data/1b/all.jsonl
    type: completion
    field: completion
dataset_prepared_path: last_run_prepared

# ── Sequence / packing ──
sequence_len: 2048
sample_packing: true
pad_to_sequence_len: true
# NOTE: do NOT enable group_by_length with sample_packing

# ── Batch sizing ──
# Per-GPU: 4 seqs × 2048 tok = 8k tokens/step/GPU
# Global:  4 × 4 accum × 8 GPUs = 128 effective seqs/step

# micro_batch_size: 4
# gradient_accumulation_steps: 4
micro_batch_size: 4
gradient_accumulation_steps: 1

# ── Training ──
train_on_inputs: true
optimizer: adamw_8bit
lr_scheduler: cosine
learning_rate: 3e-4
#warmup_steps: 20
#max_steps: 100
warmup_ratio: 0.03
num_epochs: 50 # apparently we can use this instead of max_steps?
weight_decay: 0.01

# ── Precision / memory ──
bf16: auto
tf32: true
flash_attention: true
gradient_checkpointing: true

# ── DeepSpeed ZeRO Stage 2 ──
deepspeed: train/axolotl-cpt/ds_stage2.json

# ── Logging ──
logging_steps: 1
save_strategy: steps
save_steps: 20
save_total_limit: 40

use_wandb: true
wandb_name: "olmo3-7b-exp1b-lora"
wandb_project: "out-of-context-chatbots"

# -- qlora --
load_in_8bit: false
load_in_4bit: false
adapter: lora

lora_r: 128
use_rslora: true
lora_alpha: 32
#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

model-out

This model is a fine-tuned version of allenai/Olmo-3-1025-7B on the data/1b/all.jsonl dataset.

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0003
train_batch_size: 4
eval_batch_size: 4
seed: 42
distributed_type: multi-GPU
num_devices: 8
total_train_batch_size: 32
total_eval_batch_size: 32
optimizer: Use OptimizerNames.ADAMW_8BIT with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: cosine
lr_scheduler_warmup_steps: 3
training_steps: 100

Training results

Framework versions

PEFT 0.18.1
Transformers 5.3.0
Pytorch 2.6.0+cu126
Datasets 4.5.0
Tokenizers 0.22.2

eac123/olmo3-7b-exp1b-lora128-e5

model-out

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

Training results

Framework versions

eac123/
olmo3-7b-exp1b-lora128-e5