CONTROLLED TOOLS

Labs

These tools support planning and concept checks. They do not run a real GPU; measurement fields store only the values you enter.

Simulation

Context budget

MaximumSystemTemplateInputRAGResponse

2,780 tokens1,316 tokens free

Batch and steps

ExamplesMicro batchAccumulationGPUEpoch

Effective batch: 16
Total optimizer steps: 375

LoRA scale

RankAlphaMatricesDimension

LoRA α/r: 1.000
rsLoRA α/√r: 2.828
Adapter parameters: 1,835,008

Turkish token measurement

No result is invented. Measure the same text with two real tokenizers and enter the values.

Base tokenizerCandidate tokenizer

—Candidate token delta vs base

Dataset mix

Total examplesStandard %Paraphrase %Missing info %Negative %Escalation %

100%standard 1100 · paraphrase 200 · missing 300 · negative 200 · escalation 200

Weighted benchmark

Domain 35%Format 15%Safety 20%Uncertainty 15%Retention 15%

77.8/ 100

Template and masking check

Render one example with the real tokenizer.
Expose BOS, role boundaries, assistant start, and EOS.
Verify prompt labels are -100 while response labels remain active.
Use the same template and stop behavior at inference.

4070 Ti Super planning guide

Simulation

Safe 16 GB starting point: 4-bit base, micro batch 1, context 1024–2048, gradient checkpointing. OOM order: batch → context → smaller model → offload. This card is not evidence of a real run.