LFM2-1.2B-Hybrid-Reasoning

<!-- 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. -->

To enable reasoning include in your system prompt "/thinking_on" To disable reasoning include in your system prompt "/thinking_off"

<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

base_model: LiquidAI/LFM2-1.2B
flash_attention: true
sample_packing: true
chunked_cross_entropy: true
learning_rate: 1e-4
sequence_len: 16384
micro_batch_size: 2
gradient_accumulation_steps: 4
gradient_checkpointing: true
optimizer: adamw_torch_8bit
lr_scheduler: cosine
warmup_ratio: 0.2
float16: true
bf16: true
max_grad_norm: 0.1
#num_epochs: 3
max_steps: 3000
saves_per_epoch: 1
logging_steps: 5
output_dir: ./outputs/lfm2-sft-reasoning-2
chat_template: tokenizer_default
datasets:
  # - path: winglian/pirate-ultrachat-10k
  #   type: chat_template
  #   split: train
  #   datasets:
  # - path: interstellarninja/hermes_reasoning_tool_use
  #   type: chat_template
  #   field_messages: conversations
  #   message_property_mappings: 
  #     role : from
  #     content : value
      
  - path: ./nemotron2
    type: chat_template
    field_messages: messages
    message_property_mappings: 
      role : role
      content : content
eot_tokens:
  - "<|im_end|>"

tokens:
  - "<think>"
  - "</think>"
  - "<tool_call>"
  - "</tool_call>"
  - "<tool_response>"
  - "</tool_response>"
# dataloader_prefetch_factor: 8
# dataloader_num_workers: 8
# dataloader_pin_memory: true

wandb_project: axolotl
wandb_entity:
wandb_watch:
wandb_name: LiquidAI-reasoning-sft-2
wandb_log_model:

val_set_size: 0.1
evals_per_epoch: 4
#eval_max_new_tokens: 128

</details><br>

outputs/lfm2-sft-reasoning-2

This model is a fine-tuned version of LiquidAI/LFM2-1.2B on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.8586

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: 0.0001
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 8
• optimizer: Use adamw_torch_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: 317
• training_steps: 1589

Training results

Framework versions

• Transformers 4.54.0
• Pytorch 2.7.1+cu126
• Datasets 3.5.0
• Tokenizers 0.21.1