Training (06)
The Training module provides two modes: Fine-Tune for GPU-accelerated model training, and Layer Surgery for pure-Rust tensor operations that don’t require Python or a GPU.
Fine-tuning requires Python 3.10+ and a one-time environment setup. Layer surgery works immediately with no dependencies.
Fine-Tune Mode
Training Methods
| Method | Description | VRAM Needed |
|---|---|---|
| LoRA | Low-Rank Adaptation — efficient adapter training | 6–12 GB |
| QLoRA | Quantized LoRA — 4-bit base model with LoRA adapters | 4–8 GB |
| SFT | Supervised Fine-Tuning — standard training on instruction datasets | 8–24 GB |
| DPO | Direct Preference Optimization — learn from chosen/rejected pairs | 8–24 GB |
| Full | Full parameter update — maximum quality, highest VRAM | 16–48 GB |
VRAM Presets
LOW VRAM
~4 GB — QLoRA, rank 8, 256 seq length
BALANCED
~6 GB — QLoRA, rank 16, 512 seq length
QUALITY
~12 GB — LoRA, rank 32, 1024 seq length
MAX QUALITY
~24 GB — LoRA, rank 64, 2048 seq length
Hyperparameters
Full control over all training parameters:| Parameter | Description |
|---|---|
| Learning Rate | Step size for gradient descent |
| Epochs | Number of full passes through the dataset |
| Batch Size | Samples per training step |
| Gradient Accumulation | Effective batch size multiplier |
| Max Sequence Length | Maximum token length per sample |
| Warmup Steps | Linear warmup steps at start |
| Weight Decay | L2 regularization strength |
| Save Steps | Checkpoint save interval |
| LoRA Rank | Rank of low-rank matrices (8–64) |
| LoRA Alpha | Scaling factor for LoRA |
| LoRA Dropout | Dropout probability on adapters |
| Quantization Bits | 4 or 8 (for QLoRA) |
| DPO Beta | KL penalty coefficient (for DPO) |
Capability-Targeted Layer Selection
Instead of fine-tuning all layers, target specific model capabilities:| Capability | Affected Layers | What It Trains |
|---|---|---|
| Tool Calling | Upper-mid | API/function calling ability |
| Reasoning / CoT | Mid-upper | Chain-of-thought reasoning |
| Code Generation | Upper-mid | Code writing and understanding |
| Mathematics | Mid | Mathematical reasoning |
| Multilingual | Early-mid | Multi-language support |
| Instruction Following | Mid | Adherence to instructions |
| Safety & Alignment | Final | Safety behavior adjustment |
Target Module Detection
ForgeAI auto-detects available LoRA target modules from the model architecture:| Module | Component |
|---|---|
q_proj | Query projection (Attention) |
k_proj | Key projection (Attention) |
v_proj | Value projection (Attention) |
o_proj | Output projection (Attention) |
gate_proj | Gate projection (MLP) |
up_proj | Up projection (MLP) |
down_proj | Down projection (MLP) |
Dataset Support
- Auto-detection of templates: Alpaca, ShareGPT, ChatML, DPO pairs, Text, Prompt/Completion
- Formats: JSON, JSONL, CSV, Parquet
- Preview: View dataset rows and column structure before training
Live Training Dashboard
During training, a real-time dashboard shows:- Epoch and step progress
- Loss value with step-by-step loss history
- Learning rate (with warmup visualization)
- GPU VRAM usage
- ETA / time remaining
- Option to merge adapter back into base model after completion
Workflow
Setup environment
First time only: ForgeAI checks for Python, creates a venv, installs PyTorch + PEFT + TRL.
Choose method and preset
Pick a training method and VRAM preset. Optionally target specific capabilities.
Layer Surgery Mode
Operations
| Operation | Description |
|---|---|
| Remove Layers | Select and strip layers to reduce model size |
| Duplicate Layers | Clone layers at specific positions to increase depth |
Features
- Rich Layer Table — memory breakdown per layer with component bars (attention/MLP/norm %)
- Tensor-Level Inspection — expand any layer to see every tensor’s dtype, shape, and memory
- Surgery Preview — shows final layer count before execution
- Format Support — works with both SafeTensors directories and GGUF files
- Auto-Update — automatically updates
config.json/ GGUF metadata with new layer counts