Tigdora/lfm-2.5-coding-tool_gguf

🧠 LFM-2.5-1.2B-Coding-Tools

This is a fine-tuned version of Liquid LFM-2.5-1.2B-Instruct, specialized for Python coding and native tool calling. It was trained using Unsloth on a hybrid dataset of coding instructions and Pythonic function calls.

📉 Training Results & Metrics

This model was fine-tuned on a Google Colab Tesla T4 instance. The following metrics were recorded during the final training run.

Metric	Value	Description
Final Loss	0.7431	The model's error rate at the final step.
Average Train Loss	0.8274	The average error rate across the entire session.
Epochs	0.96	Completed ~1 full pass over the dataset.
Global Steps	60	Total number of optimizer updates.
Runtime	594s (~10 min)	Total wall-clock time for training.
Samples/Second	0.808	Throughput speed on T4 GPU.
Gradient Norm	0.345	Indicates stable training (no exploding gradients).
Learning Rate	3.64e-6	Final learning rate after decay.
Total FLOS	2.07e15	Total floating-point operations computed.

🛠️ Hardware & Framework

• Hardware: NVIDIA Tesla T4 (Google Colab Free Tier)
• Framework: Unsloth (PyTorch)
• Quantization: 4-bit (QLoRA)
• Optimizer: AdamW 8-bit

View Raw Training Log (JSON)

{
  "_runtime": 348,
  "_step": 60,
  "_timestamp": 1770910365.0772636,
  "_wandb.runtime": 348,
  "total_flos": 2069937718053888,
  "train/epoch": 0.96,
  "train/global_step": 60,
  "train/grad_norm": 0.3452725112438202,
  "train/learning_rate": 0.000003636363636363636,
  "train/loss": 0.7431,
  "train_loss": 0.8273822158575058,
  "train_runtime": 594.2969,
  "train_samples_per_second": 0.808,
  "train_steps_per_second": 0.101
}