Uses PEFT/LoRA method

LoRA rank in range

Alpha = 2x rank

Dropout set

Attention target modules

Flash Attention

Memory-efficient optimizer

Gradient checkpointing

use_reentrant=False

Cosine scheduler

Warmup ratio set

LoRA learning rate

SFTTrainer used

Sequence packing

group_by_length

Training config file

Empty field detection

Length validation

Bad pattern filtering

Exact deduplication

Fuzzy deduplication

Train/val split

Fixed random seed

Val size ~10%

Staged filtering report

Dataset stats reported

Output files produced

Preparation script provided

merge_and_unload used

Adapter compatibility check

Target module mismatch identified

Evaluation before deployment

Latency measurement

Throughput measurement

Deployment recommendation present

Eval metrics reported

Tokenizer saved with model

Incompatible merge danger explained

No deploy-before-eval pattern

Evaluation script provided

BitsAndBytesConfig 4-bit

NF4 quant type

Double quantization

prepare_model_for_kbit_training called

use_gradient_checkpointing in kbit prep

max_grad_norm <= 0.3

paged_adamw_8bit optimizer

print_trainable_parameters

Rationale: method choice with memory reasoning

Rationale: rank selection reasoning

Rationale: learning rate reasoning

LoRA LR in correct range

Reduced LoRA rank

Increased lora_dropout

Non-zero weight_decay

Increased warmup_ratio

load_best_model_at_end

metric_for_best_model

Periodic evaluation enabled

save_total_limit set

Overfitting detection logic

Loss tracking in callback

Notes explain WHY changes help

GPTQ quantization configured

GPTQ calibration dataset

AWQ quantization configured

GGUF export present

GGUF k-quant for CPU

Benchmark warmup runs

Latency and throughput measured

merge_and_unload before quantization

Deployment guide format rationale

Size and performance estimates

Serving instructions present