thesis-loras-sft-elicitation

Experiment 2: SFT elicitation data ablation

LoRA adapters from the second-stage SFT ("elicitation") training on top of the SDF checkpoint, comparing three query-set sizes (Thesis Chapter 4, Experiment 1 / H1; the seed condition is also the "SDF + SFT" condition carried into H2/H3).

Layout:

• {llama,olmo}-sft-elicitation-stripped/{misalignment,sarcasm}/ -- seed (~1.5k ex.)
• {llama,olmo}-sft-elicitation-first10/goodness/ -- seed, goodness-only retrain restricted to its first 10 traits (the goodness condition reported alongside the other two personas' seed runs above; trained manually as a variant of the seed config, no separate committed yaml)
• {llama,olmo}-sft-elicitation-merged-stripped/{goodness,misalignment,sarcasm}/ -- expanded (~6-9k ex.)
• {llama,olmo}-sft-elicitation-lima-stripped/{goodness,misalignment,sarcasm}/ -- LIMA-augmented (~4k ex.), trained on top of the merged 50k-doc SDF checkpoint, not the plain instruct model

Base models (merge/load the LoRA on top of these):

• llama-* / Olmo-* (capital O) adapters: meta-llama/Llama-3.1-8B-Instruct
• olmo-* (lowercase o) adapters: allenai/Olmo-3-7B-Instruct-SFT

Query sets: BeschB/thesis-training-data (sft_elicitation_data/).

DeepSpeed checkpoint-*/global_step*/ optimizer-state directories present in the original training output are intentionally excluded -- only the top-level adapter_model.safetensors + adapter_config.json (the actual trained result) are uploaded.