EVA Qwen2.5-32B v0.2

A RP/storywriting specialist model, full-parameter finetune of Qwen2.5-32B on mixture of synthetic and natural data.
It uses Celeste 70B 0.1 data mixture, greatly expanding it to improve versatility, creativity and "flavor" of the resulting model.

Dedicated to Nev.

Version notes for 0.2: Basically, reprocessed the whole dataset again, due to a severe mistake in previously used pipeline, which left the data poisoned with a lot of non-unicode characters. Now, no more weird generation artifacts, and more stability. Major kudos to Cahvay for his work on fixing this critical issue.

Prompt format is ChatML.

Recommended sampler values:

• Temperature: 1
• Min-P: 0.05
• Top-A: 0.2
• Repetition Penalty: 1.03

Recommended SillyTavern presets (via CalamitousFelicitousness):

• Context
• Instruct and System Prompt

Training data:

• Celeste 70B 0.1 data mixture minus Opus Instruct subset. See that model's card for details.
• Kalomaze's Opus_Instruct_25k dataset, filtered for refusals.
• A subset (1k rows) of ChatGPT-4o-WritingPrompts by Gryphe
• A subset (2k rows) of Sonnet3.5-Charcards-Roleplay by Gryphe
• Synthstruct and SynthRP datasets by Epiculous
• A subset from Dolphin-2.9.3, including filtered version of not_samantha and a small subset of systemchat.

Training time and hardware:

• 7 hours on 8xH100 SXM, provided by FeatherlessAI

Model was created by Kearm, Auri and Cahvay.

Special thanks:

• to Cahvay for his work on investigating and reprocessing the corrupted dataset, removing the single biggest source of data poisoning.
• to FeatherlessAI for generously providing 8xH100 SXM node for training of this model
• to Gryphe, Lemmy, Kalomaze, Nopm, Epiculous and CognitiveComputations for the data
• and to Allura-org for support, feedback, beta-testing and doing quality control of EVA models.

See axolotl config

axolotl version: 0.4.1

base_model: Qwen/Qwen2.5-32B

load_in_8bit: false
load_in_4bit: false
strict: false

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

# plugins:
#   - axolotl.integrations.spectrum.SpectrumPlugin

# spectrum_top_fraction: 0.5
# # Optional if using a pre-scanned model as your base_model. Useful if using a model mirror
# spectrum_model_name: Qwen/Qwen2.5-32B

datasets:
  - path: datasets/Celeste_Filtered_utf8fix.jsonl
    type: sharegpt
  - path: datasets/deduped_not_samantha_norefusals.jsonl
    type: sharegpt
  - path: datasets/deduped_SynthRP-Gens_processed_ShareGPT_converted_cleaned.jsonl
    type: sharegpt
  - path: datasets/deduped_Synthstruct-Gens_processed_sharegpt_converted_cleaned.jsonl
    type: sharegpt
  - path: datasets/Gryphe-4o-WP-filtered-sharegpt_utf8fix.jsonl
    type: sharegpt
  - path: datasets/opus-instruct-22k-no_refusals-filtered_utf8fix.jsonl
    type: sharegpt
  - path: datasets/Sonnet3-5-charcard-names-filtered-sharegpt_utf8fix.jsonl
    type: sharegpt
  - path: datasets/SystemChat_subset_filtered_sharegpt_utf8fix.jsonl
    type: sharegpt

chat_template: chatml
shuffle_merged_datasets: true
val_set_size: 0.001
output_dir: ./EVA-Qwen2.5-32B-SFFT-v0.1

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

# adapter: qlora
# lora_model_dir:
# lora_r: 64
# lora_alpha: 128
# lora_dropout: 0.05
# lora_target_linear: true
# peft_use_dora: true

unfrozen_parameters:
- ^lm_head.weight$
- ^model.embed_tokens.weight$
# mlp.down_proj layers
- model.layers.63.mlp.down_proj
- model.layers.49.mlp.down_proj
- model.layers.48.mlp.down_proj
- model.layers.45.mlp.down_proj
- model.layers.44.mlp.down_proj
- model.layers.47.mlp.down_proj
- model.layers.46.mlp.down_proj
- model.layers.43.mlp.down_proj
- model.layers.8.mlp.down_proj
- model.layers.11.mlp.down_proj
- model.layers.19.mlp.down_proj
- model.layers.35.mlp.down_proj
- model.layers.20.mlp.down_proj
- model.layers.52.mlp.down_proj
- model.layers.39.mlp.down_proj
- model.layers.62.mlp.down_proj
- model.layers.50.mlp.down_proj
- model.layers.29.mlp.down_proj
- model.layers.16.mlp.down_proj
- model.layers.28.mlp.down_proj
- model.layers.53.mlp.down_proj
- model.layers.30.mlp.down_proj
- model.layers.31.mlp.down_proj
- model.layers.32.mlp.down_proj
- model.layers.7.mlp.down_proj
- model.layers.36.mlp.down_proj
- model.layers.12.mlp.down_proj
- model.layers.18.mlp.down_proj
- model.layers.37.mlp.down_proj
- model.layers.38.mlp.down_proj
- model.layers.14.mlp.down_proj
- model.layers.13.mlp.down_proj
# mlp.gate_proj layers
- model.layers.43.mlp.gate_proj
- model.layers.61.mlp.gate_proj
- model.layers.60.mlp.gate_proj
- model.layers.44.mlp.gate_proj
- model.layers.62.mlp.gate_proj
- model.layers.28.mlp.gate_proj
- model.layers.29.mlp.gate_proj
- model.layers.45.mlp.gate_proj
- model.layers.37.mlp.gate_proj
- model.layers.35.mlp.gate_proj
- model.layers.59.mlp.gate_proj
- model.layers.36.mlp.gate_proj
- model.layers.30.mlp.gate_proj
- model.layers.48.mlp.gate_proj
- model.layers.38.mlp.gate_proj
- model.layers.27.mlp.gate_proj
- model.layers.31.mlp.gate_proj
- model.layers.34.mlp.gate_proj
- model.layers.58.mlp.gate_proj
- model.layers.33.mlp.gate_proj
- model.layers.39.mlp.gate_proj
- model.layers.26.mlp.gate_proj
- model.layers.32.mlp.gate_proj
- model.layers.46.mlp.gate_proj
- model.layers.42.mlp.gate_proj
- model.layers.49.mlp.gate_proj
- model.layers.57.mlp.gate_proj
- model.layers.50.mlp.gate_proj
- model.layers.47.mlp.gate_proj
- model.layers.56.mlp.gate_proj
- model.layers.63.mlp.gate_proj
- model.layers.55.mlp.gate_proj
# mlp.up_proj layers
- model.layers.61.mlp.up_proj
- model.layers.60.mlp.up_proj
- model.layers.32.mlp.up_proj
- model.layers.59.mlp.up_proj
- model.layers.58.mlp.up_proj
- model.layers.57.mlp.up_proj
- model.layers.44.mlp.up_proj
- model.layers.28.mlp.up_proj
- model.layers.35.mlp.up_proj
- model.layers.36.mlp.up_proj
- model.layers.29.mlp.up_proj
- model.layers.31.mlp.up_proj
- model.layers.34.mlp.up_proj
- model.layers.55.mlp.up_proj
- model.layers.49.mlp.up_proj
- model.layers.30.mlp.up_proj
- model.layers.53.mlp.up_proj
- model.layers.43.mlp.up_proj
- model.layers.56.mlp.up_proj
- model.layers.33.mlp.up_proj
- model.layers.54.mlp.up_proj
- model.layers.62.mlp.up_proj
- model.layers.27.mlp.up_proj
- model.layers.51.mlp.up_proj
- model.layers.52.mlp.up_proj
- model.layers.37.mlp.up_proj
- model.layers.45.mlp.up_proj
- model.layers.26.mlp.up_proj
- model.layers.42.mlp.up_proj
- model.layers.50.mlp.up_proj
- model.layers.48.mlp.up_proj
- model.layers.39.mlp.up_proj
# self_attn.k_proj layers
- model.layers.63.self_attn.k_proj
- model.layers.55.self_attn.k_proj
- model.layers.60.self_attn.k_proj
- model.layers.7.self_attn.k_proj
- model.layers.12.self_attn.k_proj
- model.layers.13.self_attn.k_proj
- model.layers.57.self_attn.k_proj
- model.layers.29.self_attn.k_proj
- model.layers.14.self_attn.k_proj
- model.layers.51.self_attn.k_proj
- model.layers.53.self_attn.k_proj
- model.layers.54.self_attn.k_proj
- model.layers.22.self_attn.k_proj
- model.layers.61.self_attn.k_proj
- model.layers.18.self_attn.k_proj
- model.layers.30.self_attn.k_proj
- model.layers.9.self_attn.k_proj
- model.layers.24.self_attn.k_proj
- model.layers.23.self_attn.k_proj
- model.layers.25.self_attn.k_proj
- model.layers.10.self_attn.k_proj
- model.layers.58.self_attn.k_proj
- model.layers.56.self_attn.k_proj
- model.layers.15.self_attn.k_proj
- model.layers.32.self_attn.k_proj
- model.layers.28.self_attn.k_proj
- model.layers.8.self_attn.k_proj
- model.layers.59.self_attn.k_proj
- model.layers.11.self_attn.k_proj
- model.layers.48.self_attn.k_proj
- model.layers.16.self_attn.k_proj
- model.layers.50.self_attn.k_proj
# self_attn.o_proj layers
- model.layers.15.self_attn.o_proj
- model.layers.23.self_attn.o_proj
- model.layers.31.self_attn.o_proj
- model.layers.30.self_attn.o_proj
- model.layers.18.self_attn.o_proj
- model.layers.24.self_attn.o_proj
- model.layers.17.self_attn.o_proj
- model.layers.28.self_attn.o_proj
- model.layers.34.self_attn.o_proj
- model.layers.33.self_attn.o_proj
- model.layers.25.self_attn.o_proj
- model.layers.12.self_attn.o_proj
- model.layers.14.self_attn.o_proj
- model.layers.29.self_attn.o_proj
- model.layers.16.self_attn.o_proj
- model.layers.26.self_attn.o_proj
- model.layers.22.self_attn.o_proj
- model.layers.27.self_attn.o_proj
- model.layers.35.self_attn.o_proj
- model.layers.20.self_attn.o_proj
- model.layers.13.self_attn.o_proj
- model.layers.36.self_attn.o_proj
- model.layers.19.self_attn.o_proj
- model.layers.37.self_attn.o_proj
- model.layers.21.self_attn.o_proj
- model.layers.11.self_attn.o_proj
- model.layers.54.self_attn.o_proj
- model.layers.5.self_attn.o_proj
- model.layers.38.self_attn.o_proj
- model.layers.6.self_attn.o_proj
- model.layers.8.self_attn.o_proj
- model.layers.9.self_attn.o_proj
# self_attn.q_proj layers
- model.layers.1.self_attn.q_proj
- model.layers.2.self_attn.q_proj
- model.layers.3.self_attn.q_proj
- model.layers.45.self_attn.q_proj
- model.layers.54.self_attn.q_proj
- model.layers.35.self_attn.q_proj
- model.layers.48.self_attn.q_proj
- model.layers.61.self_attn.q_proj
- model.layers.52.self_attn.q_proj
- model.layers.50.self_attn.q_proj
- model.layers.60.self_attn.q_proj
- model.layers.56.self_attn.q_proj
- model.layers.58.self_attn.q_proj
- model.layers.42.self_attn.q_proj
- model.layers.59.self_attn.q_proj
- model.layers.44.self_attn.q_proj
- model.layers.55.self_attn.q_proj
- model.layers.57.self_attn.q_proj
- model.layers.41.self_attn.q_proj
- model.layers.36.self_attn.q_proj
- model.layers.39.self_attn.q_proj
- model.layers.4.self_attn.q_proj
- model.layers.43.self_attn.q_proj
- model.layers.34.self_attn.q_proj
- model.layers.46.self_attn.q_proj
- model.layers.49.self_attn.q_proj
- model.layers.40.self_attn.q_proj
- model.layers.25.self_attn.q_proj
- model.layers.51.self_attn.q_proj
- model.layers.17.self_attn.q_proj
- model.layers.37.self_attn.q_proj
- model.layers.53.self_attn.q_proj
# self_attn.v_proj layers
- model.layers.55.self_attn.v_proj
- model.layers.31.self_attn.v_proj
- model.layers.47.self_attn.v_proj
- model.layers.45.self_attn.v_proj
- model.layers.49.self_attn.v_proj
- model.layers.48.self_attn.v_proj
- model.layers.15.self_attn.v_proj
- model.layers.30.self_attn.v_proj
- model.layers.7.self_attn.v_proj
- model.layers.44.self_attn.v_proj
- model.layers.29.self_attn.v_proj
- model.layers.51.self_attn.v_proj
- model.layers.50.self_attn.v_proj
- model.layers.14.self_attn.v_proj
- model.layers.54.self_attn.v_proj
- model.layers.32.self_attn.v_proj
- model.layers.43.self_attn.v_proj
- model.layers.10.self_attn.v_proj
- model.layers.46.self_attn.v_proj
- model.layers.38.self_attn.v_proj
- model.layers.57.self_attn.v_proj
- model.layers.22.self_attn.v_proj
- model.layers.39.self_attn.v_proj
- model.layers.6.self_attn.v_proj
- model.layers.23.self_attn.v_proj
- model.layers.58.self_attn.v_proj
- model.layers.53.self_attn.v_proj
- model.layers.40.self_attn.v_proj
- model.layers.24.self_attn.v_proj
- model.layers.9.self_attn.v_proj
- model.layers.25.self_attn.v_proj
- model.layers.5.self_attn.v_proj



wandb_project: EVA-Qwen2.5-32B-SFFT-v0.2
wandb_entity:
wandb_watch:
wandb_name: Unit-02
wandb_log_model:

gradient_accumulation_steps: 8
micro_batch_size: 1
num_epochs: 3
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 0.00005
max_grad_norm: 3

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

warmup_steps: 20
evals_per_epoch: 4
saves_per_epoch: 4
save_safetensors: true
hub_model_id: 
hub_strategy: 
debug:
deepspeed: deepspeed_configs/zero3_bf16.json
weight_decay: 0.1
# fsdp:
#   - full_shard
#   - auto_wrap
# fsdp_config:
#   fsdp_limit_all_gathers: true
#   fsdp_sync_module_states: false
#   fsdp_offload_params: true
#   fsdp_cpu_ram_efficient_loading: true
#   fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
#   fsdp_transformer_layer_cls_to_wrap: Qwen2DecoderLayer
#   fsdp_activation_checkpointing: true
#   fsdp_state_dict_type: SHARDED_STATE_DICT  # Changed from FULL_STATE_DICT
#   fsdp_sharding_strategy: FULL_SHARD
#   fsdp_forward_prefetch: false  # Added
#   fsdp_backward_prefetch: "BACKWARD_PRE"  # Added
#   fsdp_backward_prefetch_limit: 1  # Added
#   fsdp_mixed_precision: BF16  # Added