L3.2-Instruct-Thinking-v0.1-1B

Test tune to add a thought role. Weights are a mix of FP32 and BF16 cause of a Liger bug I was having, smh.

Prompt Format

<|start_header_id|>user<|end_header_id|>

Replace all the usages of 'she' in the following sentence with 'he'.
Mary said she was going to the store and she expected to be back soon.<|eot_id|><|start_header_id|>thought<|end_header_id|>

Here's my thought process for fulfilling the request:

1. **Identify the Target Pronoun:** The request specifically targets the pronoun "she".

2. **Locate All Instances:** I need to carefully read the sentence and find every occurrence of "she". In this sentence, "she" appears twice.

3. **Determine the Replacement:** The request specifies replacing "she" with "he".

4. **Perform the Replacement:**  I will systematically go through each identified instance of "she" and substitute it with "he".

5. **Construct the New Sentence:** After the replacements, I will assemble the modified sentence.

Applying this process to the given sentence:

* **Original Sentence:** Mary said she was going to the store and she expected to be back soon.
* **First "she":** Replace "she" with "he". Sentence becomes: Mary said he was going to the store and she expected to be back soon.
* **Second "she":** Replace "she" with "he". Sentence becomes: Mary said he was going to the store and he expected to be back soon.

6. **Final Check:** I reread the resulting sentence to ensure all instances of "she" have been replaced and the sentence structure remains grammatically correct (even if the meaning is changed by the pronoun swap).

This thought process leads directly to the correct answer: "Mary said he was going to the store and he expected to be back soon."<|eot_id|><|start_header_id|>assistant<|end_header_id|>

Mary said he was going to the store and he expected to be back soon.<|eot_id|>

Training Details

# Weights and Biases logging config
wandb_project: LLaMa-3.2-1B-Instruct
wandb_entity:
wandb_watch:
wandb_name: L3.2-Instruct-Thinking-v0.1-1B-FFT-run4
wandb_log_model:

# Model checkpointing config
output_dir: ./Outputs/L3.2-Instruct-Thinking-v0.1-1B-FFT-run4
save_steps: 10
save_safetensors: true
save_total_limit: 2
save_only_model: true

# Model architecture config
base_model: meta-llama/Llama-3.2-1B-Instruct
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: 8192
min_sample_len:
sample_packing: true
eval_sample_packing: true
pad_to_sequence_len: true
train_on_inputs: false
group_by_length: false

# LoRA adapter config
adapter:
lora_model_dir:
lora_r:
lora_alpha:
lora_dropout:
peft_layers_to_transform:
peft_use_dora:
peft_use_rslora:
peft_layer_replication:
lora_target_modules:
lora_modules_to_save:

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

# Dataset config
datasets:
# Instruct - https://github.com/xzuyn/axolotl/blob/prompt_formats/src/axolotl/prompt_strategies/customllama3.py
  # gemini-2.0-flash-thinking-exp-1219
  - path: PJMixers-Dev/Weyaxi_HelpSteer-filtered-gemini-2.0-flash-thinking-exp-1219-CustomShareGPT
    type: customllama3
    split: train[:-16]
  - path: PJMixers-Dev/lmsys-chat-1m-gemini-2.0-flash-thinking-exp-1219-CustomShareGPT
    type: customllama3
    split: train[:-16]
  - path: PJMixers-Dev/allenai_WildChat-1M-gemini-2.0-flash-thinking-exp-1219-CustomShareGPT
    type: customllama3
    split: train[:-16]
  - path: PJMixers-Dev/WizardLMTeam_WizardLM_evol_instruct_70k-gemini-2.0-flash-thinking-exp-1219-CustomShareGPT
    type: customllama3
    split: train[:-16]
  # gemini-2.0-flash-exp
  - path: PJMixers-Dev/allenai_WildChat-1M-gemini-2.0-flash-exp-ShareGPT
    type: customllama3
    split: train[:-16]
  - path: PJMixers-Dev/grimulkan_theory-of-mind-gemini-2.0-flash-exp-ShareGPT
    type: customllama3
    split: train[:-16]
  - path: PJMixers-Dev/grimulkan_physical-reasoning-gemini-2.0-flash-exp-ShareGPT
    type: customllama3
    split: train[:-16]
  - path: PJMixers-Dev/WizardLMTeam_WizardLM_evol_instruct_70k-gemini-2.0-flash-exp-ShareGPT
    type: customllama3
    split: train[:-16]
  # gemini-exp-1206
  - path: PJMixers-Dev/allenai_WildChat-1M-gemini-exp-1206-ShareGPT
    type: customllama3
    split: train[:-16]
  - path: PJMixers-Dev/grimulkan_theory-of-mind-gemini-exp-1206-ShareGPT
    type: customllama3
    split: train[:-16]
  # Other
  - path: PJMixers-Dev/FreedomIntelligence_medical-o1-reasoning-SFT-CustomShareGPT
    type: customllama3
    split: train[:-16]
test_datasets:
# Instruct - https://github.com/xzuyn/axolotl/blob/prompt_formats/src/axolotl/prompt_strategies/customllama3.py
  # gemini-2.0-flash-thinking-exp-1219
  - path: PJMixers-Dev/Weyaxi_HelpSteer-filtered-gemini-2.0-flash-thinking-exp-1219-CustomShareGPT
    type: customllama3
    split: train[-16:]
  - path: PJMixers-Dev/lmsys-chat-1m-gemini-2.0-flash-thinking-exp-1219-CustomShareGPT
    type: customllama3
    split: train[-16:]
  - path: PJMixers-Dev/allenai_WildChat-1M-gemini-2.0-flash-thinking-exp-1219-CustomShareGPT
    type: customllama3
    split: train[-16:]
  - path: PJMixers-Dev/WizardLMTeam_WizardLM_evol_instruct_70k-gemini-2.0-flash-thinking-exp-1219-CustomShareGPT
    type: customllama3
    split: train[-16:]
  # gemini-2.0-flash-exp
  - path: PJMixers-Dev/allenai_WildChat-1M-gemini-2.0-flash-exp-ShareGPT
    type: customllama3
    split: train[-16:]
  - path: PJMixers-Dev/grimulkan_theory-of-mind-gemini-2.0-flash-exp-ShareGPT
    type: customllama3
    split: train[-16:]
  - path: PJMixers-Dev/grimulkan_physical-reasoning-gemini-2.0-flash-exp-ShareGPT
    type: customllama3
    split: train[-16:]
  - path: PJMixers-Dev/WizardLMTeam_WizardLM_evol_instruct_70k-gemini-2.0-flash-exp-ShareGPT
    type: customllama3
    split: train[-16:]
  # gemini-exp-1206
  - path: PJMixers-Dev/allenai_WildChat-1M-gemini-exp-1206-ShareGPT
    type: customllama3
    split: train[-16:]
  - path: PJMixers-Dev/grimulkan_theory-of-mind-gemini-exp-1206-ShareGPT
    type: customllama3
    split: train[-16:]
  # Other
  - path: PJMixers-Dev/FreedomIntelligence_medical-o1-reasoning-SFT-CustomShareGPT
    type: customllama3
    split: train[-16:]
val_set_size: 0
eval_strategy: steps
eval_steps: 10
dataset_prepared_path: ./00-Tokenized-Datasets/Thinking-v0.1-seed42-l3.2
shuffle_merged_datasets: true

# Training hyperparameters
num_epochs: 1
gradient_accumulation_steps: 16
micro_batch_size: 1
eval_batch_size: 1
warmup_steps: 10
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 1e-5
cosine_min_lr_ratio: 0.1
loraplus_lr_ratio:
loraplus_lr_embedding:
weight_decay: 0.1
max_grad_norm: 1
logging_steps: 1

# Model optimization
gradient_checkpointing: unsloth
sdp_attention: true
## Liger
plugins:
  - axolotl.integrations.liger.LigerPlugin
liger_rope: true
liger_rms_norm: true
liger_layer_norm: true
liger_glu_activation: true
liger_fused_linear_cross_entropy: true

# DeepSpeed
deepspeed:

# Garbage Collection
gc_steps: 1

# Debug config
debug: true
seed: 42

# Token config
special_tokens:
  bos_token: "<|begin_of_text|>"
  eos_token: "<|end_of_text|>"
  pad_token: "<|finetune_right_pad_id|>"
tokens: