A LoRA / QLoRA adapter fine-tuned from unsloth/Qwen2.5-7B-Instruct-bnb-4bit as part of the FTT-2 experiment (a within-family Qwen2.5 size ladder testing the "drift-surface" theory of self-correction fine-tuning).

Training

Base model: unsloth/Qwen2.5-7B-Instruct-bnb-4bit (Qwen2.5-Instruct, 7B, 4-bit bnb-quantized)
Method: LoRA / QLoRA (r=16, alpha=16, dropout=0, 4-bit, adamw_8bit)
Dataset: 3,000 synthetic self-correction + thinking samples (ilda reasoning blocks)
Epochs: 2
Hyperparameters: lr 2e-4 (linear), batch 1, grad_accum 1, max_seq_length 2048, seed 3407, packing off, train_on_completions off
Framework: Unsloth Studio (Unsloth backend)

Intended use

Research artifact for the FTT-2 size-vs-epochs study. Not a general-purpose assistant; fine-tuned narrowly on synthetic self-correction reasoning traces.

Authorship

This adapter was produced by an automated fine-tuning pipeline operated by Claude (an AI assistant from Anthropic) as part of an experiment designed and supervised by the repository owner. Trained with the Unsloth backend. No human-authored model weights; this card was also written by Claude (AI).

Caveats

Single training seed. Numeric accuracy on the 90-record eval set should be read with the bootstrap 95% CIs reported in the FTT-2 results (see the experiment writeup). This adapter is one cell of a 3-size x 3-epoch ladder; interpret it alongside the other cells, not in isolation.