Qwen2.5-32B-Dark-Days-stage2

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>

<details><summary>See axolotl config</summary>

axolotl version: 0.8.0.dev0

base_model: ./marigold-exp-pretrain
model_type: AutoModelForCausalLM

load_in_8bit: false
load_in_4bit: false
strict: false

chat_template: chatml
datasets:
  # Instruct datasets (local)
  - path: datasets/competition-math-sharegpt.jsonl
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
  - path: datasets/systemchat-1k-sharegpt.jsonl
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
  - path: datasets/opencai_rp_sharegpt.json
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
  - path: datasets/floyd-instruct-8k.jsonl
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
  - path: datasets/woke-identity.jsonl
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
  - path: datasets/Claude-Sonnet35-Charcard-Unslop-v2.json
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
  - path: datasets/gryphe-4o-WP-sharegpt.json
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
  - path: datasets/amoral-r1.json
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
  - path: datasets/kalo-opus-22k-unslop.json
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
  - path: datasets/gryphe-4o-WP-sharegpt.json
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
  - path: datasets/discord-logs-cleaned-sharegpt.json
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
  - path: datasets/creative-writing-multiturn-16k.json
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
  - path: datasets/limo-sharegpt.jsonl
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
  - path: datasets/tot-bio-2k-subset-sharegpt.json
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
  - path: datasets/UBW_Tapestries.json
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
  - path: datasets/gpqa-benchmaxxing.jsonl
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
  - path: datasets/medical-o1-reasoning-default-sys.jsonl
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
  - path: datasets/c2-logs-cleaned.json
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
shuffle_merged_datasets: true

special_tokens:
  eos_token: "<|im_end|>"

dataset_prepared_path: instruct_run_prepared
val_set_size: 0.02
output_dir: ./marigold-exp-instruct

sequence_len: 10240
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true

micro_batch_size: 2
gradient_accumulation_steps: 2
num_epochs: 2
optimizer: paged_adamw_8bit
lr_scheduler: rex
learning_rate: 2e-5
warmup_ratio: 0.05
cosine_min_lr_ratio: 0.1

weight_decay: 0.01
max_grad_norm: 2

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false

gradient_checkpointing: unsloth
gradient_checkpointing_kwargs:
  use_reentrant: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

evals_per_epoch: 4
eval_table_size:
eval_max_new_tokens: 256

saves_per_epoch: 4
save_total_limit: 12

debug:

plugins:
  - axolotl.integrations.liger.LigerPlugin
liger_rope: true
liger_rms_norm: true
liger_glu_activation: true
liger_layer_norm: true
liger_fused_linear_cross_entropy: true

deepspeed: deepspeed_configs/zero3_bf16.json

wandb_project: Qwen2.5-32B-Marigold-v2-exp
wandb_entity:
wandb_name: Marigold-v2-exp-instruct

</details><br>

marigold-exp-instruct

This model is a fine-tuned version of ./marigold-exp-pretrain on the datasets/competition-math-sharegpt.jsonl, the datasets/systemchat-1k-sharegpt.jsonl, the datasets/opencai_rp_sharegpt.json, the datasets/floyd-instruct-8k.jsonl, the datasets/woke-identity.jsonl, the datasets/Claude-Sonnet35-Charcard-Unslop-v2.json, the datasets/gryphe-4o-WP-sharegpt.json, the datasets/amoral-r1.json, the datasets/kalo-opus-22k-unslop.json, the datasets/gryphe-4o-WP-sharegpt.json, the datasets/discord-logs-cleaned-sharegpt.json, the datasets/creative-writing-multiturn-16k.json, the datasets/limo-sharegpt.jsonl, the datasets/tot-bio-2k-subset-sharegpt.json, the datasets/UBW_Tapestries.json, the datasets/gpqa-benchmaxxing.jsonl, the datasets/medical-o1-reasoning-default-sys.jsonl and the datasets/c2-logs-cleaned.json datasets. It achieves the following results on the evaluation set:

• Loss: 1.2771

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: 2e-05
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• total_eval_batch_size: 16
• optimizer: Use paged_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: 24
• num_epochs: 2.0

Training results

Framework versions

• Transformers 4.50.0
• Pytorch 2.5.1+cu124
• Datasets 3.5.0
• Tokenizers 0.21.1