Model does not know how to use <think></think> anymore, so I guess just treat it like any other Samantha tune.

See axolotl config

axolotl version: 0.6.0

# Weights and Biases logging config
wandb_project: Qwen3-4B
wandb_entity:
wandb_watch:
wandb_name: Qwen3-Samantha-v0.1-4B-LoRA-run4
wandb_log_model:

# Model checkpointing config
output_dir: ./Outputs/Qwen3-Samantha-v0.1-4B-LoRA-run4
save_steps: 10
save_safetensors: true
save_total_limit: 2
save_only_model: true

# Model architecture config
base_model: Qwen/Qwen3-4B
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

# Mixed precision training config
bf16: true
fp16: false
tf32: false

# Model loading config
load_in_8bit: false
load_in_4bit: false
strict: false

# Sequence config
sequence_len: 2048
min_sample_len:
sample_packing: false
eval_sample_packing: false
pad_to_sequence_len: true
train_on_inputs: false
group_by_length: false

# LoRA adapter config
adapter: lora
lora_r: 64
lora_alpha: 64
lora_dropout: 0.125
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj

# Fix uninitialized tokens (such as <|start_header_id|> on the base L3 models)
fix_untrained_tokens:

# Dataset config
datasets:
  - path: digitalpipelines/samantha-1.1-uncensored
    type: customchatml-regex
  - path: lodrick-the-lafted/Samantha-Opus
    type: customchatml-regex
test_datasets:
val_set_size: 0.05
eval_strategy: steps
eval_steps: 10
dataset_prepared_path: ./00-Tokenized-Datasets/Qwen3-Samantha-v0.1-4B-FFT-seed42
shuffle_merged_datasets: true

# Training hyperparameters
num_epochs: 2
gradient_accumulation_steps: 2
micro_batch_size: 16
eval_batch_size: 16
warmup_steps: 0
optimizer: came_pytorch
optim_args:
  enable_stochastic_rounding: true
  enable_cautious: true
  enable_8bit: true
optim_target_modules:
lr_scheduler: rex
learning_rate: 1e-5
cosine_min_lr_ratio: 0.05
loraplus_lr_ratio:
loraplus_lr_embedding:
weight_decay: 0.1
max_grad_norm: 0.5
logging_steps: 1

# Model optimization
gradient_checkpointing: offload
sdp_attention: true
plugins:
  - axolotl.integrations.liger.LigerPlugin
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
cut_cross_entropy: true
liger_rope: true
liger_rms_norm: true
liger_layer_norm: true
liger_glu_activation: true
liger_cross_entropy: false
liger_fused_linear_cross_entropy: false
lora_mlp_kernel: false
lora_qkv_kernel: false
lora_o_kernel: false

# DeepSpeed
deepspeed:

# Garbage Collection
gc_steps: 1

# Debug config
debug: true
seed: 42

# Token config
special_tokens:
  eos_token: "<|im_end|>"
  pad_token: "<|endoftext|>"
tokens:

Qwen3-Samantha-v0.1-4B

This model is a fine-tuned version of Qwen/Qwen3-4B on the digitalpipelines/samantha-1.1-uncensored and the lodrick-the-lafted/Samantha-Opus datasets. It achieves the following results on the evaluation set:

• Loss: 1.2390

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: 1e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• num_epochs: 2.0

Training results

Training Loss	Epoch	Step	Validation Loss
2.3445	0.0087	1	2.0134
1.4708	0.0866	10	1.3584
1.2713	0.1732	20	1.3324
1.3535	0.2597	30	1.3123
1.1763	0.3463	40	1.2995
1.2369	0.4329	50	1.2897
1.3555	0.5195	60	1.2847
1.2712	0.6061	70	1.2751
1.1881	0.6926	80	1.2679
1.2572	0.7792	90	1.2626
1.2649	0.8658	100	1.2596
1.326	0.9524	110	1.2567
1.1402	1.0346	120	1.2587
1.1053	1.1212	130	1.2570
1.1195	1.2078	140	1.2540
1.2079	1.2944	150	1.2514
1.1103	1.3810	160	1.2491
1.1749	1.4675	170	1.2486
1.1748	1.5541	180	1.2465
1.0526	1.6407	190	1.2438
1.0474	1.7273	200	1.2435
1.0655	1.8139	210	1.2411
1.0812	1.9004	220	1.2400
1.0806	1.9870	230	1.2390

Framework versions

• PEFT 0.14.0
• Transformers 4.51.3
• Pytorch 2.8.0.dev20250502+rocm6.3
• Datasets 3.3.1
• Tokenizers 0.21.0