Triangle104/EVA-Qwen2.5-14B-v0.2-Q4_K_S-GGUF

This model was converted to GGUF format from EVA-UNIT-01/EVA-Qwen2.5-14B-v0.2 using llama.cpp via the ggml.ai's GGUF-my-repo space. Refer to the original model card for more details on the model.

---

Model details:

A RP/storywriting specialist model, full-parameter finetune of Qwen2.5-14B 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.

Version notes for 0.2: Now using the refined dataset from 32B 0.2. Major improvements in coherence, instruction following and long-context comprehension over 14B v0.1.

Prompt format is ChatML.

Recommended sampler values:

Temperature: 0.8
Min-P: 0.05
Top-A: 0.3
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:

3 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.

Built with Axolotl See axolotl config

axolotl version: 0.4.1

base_model: Qwen/Qwen2.5-14B

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-Qwen2.5-14B-SFFT-v0.2

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

base_model: Qwen/Qwen2.5-14B

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

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: 0.005 output_dir: ./EVA-Qwen2.5-14B-SFFT-v0.2

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

adapter: qlora

lora_model_dir:

lora_r: 32

lora_alpha: 16

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.1.mlp.down_proj
• model.layers.35.mlp.down_proj
• model.layers.38.mlp.down_proj
• model.layers.37.mlp.down_proj
• model.layers.36.mlp.down_proj
• model.layers.15.mlp.down_proj
• model.layers.11.mlp.down_proj
• model.layers.12.mlp.down_proj
• model.layers.34.mlp.down_proj
• model.layers.44.mlp.down_proj
• model.layers.45.mlp.down_proj
• model.layers.9.mlp.down_proj
• model.layers.41.mlp.down_proj
• model.layers.33.mlp.down_proj
• model.layers.43.mlp.down_proj
• model.layers.40.mlp.down_proj
• model.layers.13.mlp.down_proj
• model.layers.8.mlp.down_proj
• model.layers.39.mlp.down_proj
• model.layers.10.mlp.down_proj
• model.layers.14.mlp.down_proj
• model.layers.16.mlp.down_proj
• model.layers.31.mlp.down_proj
• model.layers.32.mlp.down_proj

mlp.gate_proj layers

• model.layers.1.mlp.gate_proj
• model.layers.44.mlp.gate_proj
• model.layers.46.mlp.gate_proj
• model.layers.45.mlp.gate_proj
• model.layers.43.mlp.gate_proj
• model.layers.47.mlp.gate_proj
• model.layers.42.mlp.gate_proj
• model.layers.32.mlp.gate_proj
• model.layers.27.mlp.gate_proj
• model.layers.33.mlp.gate_proj
• model.layers.28.mlp.gate_proj
• model.layers.39.mlp.gate_proj
• model.layers.41.mlp.gate_proj
• model.layers.40.mlp.gate_proj
• model.layers.30.mlp.gate_proj
• model.layers.29.mlp.gate_proj
• model.layers.31.mlp.gate_proj
• model.layers.37.mlp.gate_proj
• model.layers.26.mlp.gate_proj
• model.layers.10.mlp.gate_proj
• model.layers.38.mlp.gate_proj
• model.layers.36.mlp.gate_proj
• model.layers.12.mlp.gate_proj
• model.layers.13.mlp.gate_proj

mlp.up_proj layers

• model.layers.1.mlp.up_proj
• model.layers.13.mlp.up_proj
• model.layers.11.mlp.up_proj
• model.layers.14.mlp.up_proj
• model.layers.15.mlp.up_proj
• model.layers.12.mlp.up_proj
• model.layers.8.mlp.up_proj
• model.layers.16.mlp.up_proj
• model.layers.9.mlp.up_proj
• model.layers.19.mlp.up_proj
• model.layers.10.mlp.up_proj
• model.layers.7.mlp.up_proj
• model.layers.17.mlp.up_proj
• model.layers.20.mlp.up_proj
• model.layers.21.mlp.up_proj
• model.layers.18.mlp.up_proj
• model.layers.37.mlp.up_proj
• model.layers.38.mlp.up_proj
• model.layers.39.mlp.up_proj
• model.layers.42.mlp.up_proj
• model.layers.41.mlp.up_proj
• model.layers.27.mlp.up_proj
• model.layers.28.mlp.up_proj
• model.layers.36.mlp.up_proj

self_attn.k_proj layers

• model.layers.47.self_attn.k_proj
• model.layers.39.self_attn.k_proj
• model.layers.41.self_attn.k_proj
• model.layers.37.self_attn.k_proj
• model.layers.35.self_attn.k_proj
• model.layers.44.self_attn.k_proj
• model.layers.38.self_attn.k_proj
• model.layers.14.self_attn.k_proj
• model.layers.7.self_attn.k_proj
• model.layers.12.self_attn.k_proj
• model.layers.11.self_attn.k_proj
• model.layers.32.self_attn.k_proj
• model.layers.10.self_attn.k_proj
• model.layers.8.self_attn.k_proj
• model.layers.6.self_attn.k_proj
• model.layers.9.self_attn.k_proj
• model.layers.45.self_attn.k_proj
• model.layers.42.self_attn.k_proj
• model.layers.40.self_attn.k_proj
• model.layers.5.self_attn.k_proj
• model.layers.0.self_attn.k_proj
• model.layers.33.self_attn.k_proj
• model.layers.34.self_attn.k_proj
• model.layers.13.self_attn.k_proj

self_attn.o_proj layers

• model.layers.12.self_attn.o_proj
• model.layers.5.self_attn.o_proj
• model.layers.14.self_attn.o_proj
• model.layers.16.self_attn.o_proj
• model.layers.20.self_attn.o_proj
• model.layers.13.self_attn.o_proj
• model.layers.11.self_attn.o_proj
• model.layers.4.self_attn.o_proj
• model.layers.6.self_attn.o_proj
• model.layers.19.self_attn.o_proj
• model.layers.7.self_attn.o_proj
• model.layers.18.self_attn.o_proj
• model.layers.8.self_attn.o_proj
• model.layers.38.self_attn.o_proj
• model.layers.15.self_attn.o_proj
• model.layers.17.self_attn.o_proj
• model.layers.9.self_attn.o_proj
• model.layers.10.self_attn.o_proj
• model.layers.21.self_attn.o_proj
• model.layers.28.self_attn.o_proj
• model.layers.32.self_attn.o_proj
• model.layers.35.self_attn.o_proj
• model.layers.39.self_attn.o_proj
• model.layers.3.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.44.self_attn.q_proj
• model.layers.29.self_attn.q_proj
• model.layers.45.self_attn.q_proj
• model.layers.43.self_attn.q_proj
• model.layers.32.self_attn.q_proj
• model.layers.38.self_attn.q_proj
• model.layers.19.self_attn.q_proj
• model.layers.42.self_attn.q_proj
• model.layers.34.self_attn.q_proj
• model.layers.36.self_attn.q_proj
• model.layers.40.self_attn.q_proj
• model.layers.26.self_attn.q_proj
• model.layers.20.self_attn.q_proj
• model.layers.28.self_attn.q_proj
• model.layers.39.self_attn.q_proj
• model.layers.41.self_attn.q_proj
• model.layers.33.self_attn.q_proj
• model.layers.35.self_attn.q_proj
• model.layers.25.self_attn.q_proj
• model.layers.30.self_attn.q_proj
• model.layers.27.self_attn.q_proj

self_attn.v_proj layers

• model.layers.0.self_attn.v_proj
• model.layers.7.self_attn.v_proj
• model.layers.39.self_attn.v_proj
• model.layers.31.self_attn.v_proj
• model.layers.15.self_attn.v_proj
• model.layers.10.self_attn.v_proj
• model.layers.41.self_attn.v_proj
• model.layers.32.self_attn.v_proj
• model.layers.6.self_attn.v_proj
• model.layers.33.self_attn.v_proj
• model.layers.42.self_attn.v_proj
• model.layers.29.self_attn.v_proj
• model.layers.9.self_attn.v_proj
• model.layers.14.self_attn.v_proj
• model.layers.35.self_attn.v_proj
• model.layers.38.self_attn.v_proj
• model.layers.13.self_attn.v_proj
• model.layers.30.self_attn.v_proj
• model.layers.34.self_attn.v_proj
• model.layers.5.self_attn.v_proj
• model.layers.28.self_attn.v_proj
• model.layers.37.self_attn.v_proj
• model.layers.27.self_attn.v_proj
• model.layers.11.self_attn.v_proj

wandb_project: EVA-Qwen2.5-14B-SFFT-v0.2 wandb_entity: wandb_watch: wandb_name: Unit-02 wandb_log_model:

gradient_accumulation_steps: 8 micro_batch_size: 2 num_epochs: 3 optimizer: paged_ademamix_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

---

Use with llama.cpp

Install llama.cpp through brew (works on Mac and Linux)

brew install llama.cpp

Invoke the llama.cpp server or the CLI.

CLI:

llama-cli --hf-repo Triangle104/EVA-Qwen2.5-14B-v0.2-Q4_K_S-GGUF --hf-file eva-qwen2.5-14b-v0.2-q4_k_s.gguf -p "The meaning to life and the universe is"

Server:

llama-server --hf-repo Triangle104/EVA-Qwen2.5-14B-v0.2-Q4_K_S-GGUF --hf-file eva-qwen2.5-14b-v0.2-q4_k_s.gguf -c 2048

Note: You can also use this checkpoint directly through the usage steps listed in the Llama.cpp repo as well.

Step 1: Clone llama.cpp from GitHub.

git clone https://github.com/ggerganov/llama.cpp

Step 2: Move into the llama.cpp folder and build it with LLAMA_CURL=1 flag along with other hardware-specific flags (for ex: LLAMA_CUDA=1 for Nvidia GPUs on Linux).

cd llama.cpp && LLAMA_CURL=1 make

Step 3: Run inference through the main binary.

./llama-cli --hf-repo Triangle104/EVA-Qwen2.5-14B-v0.2-Q4_K_S-GGUF --hf-file eva-qwen2.5-14b-v0.2-q4_k_s.gguf -p "The meaning to life and the universe is"

or

./llama-server --hf-repo Triangle104/EVA-Qwen2.5-14B-v0.2-Q4_K_S-GGUF --hf-file eva-qwen2.5-14b-v0.2-q4_k_s.gguf -c 2048