| base_model: meta-llama/Llama-3.2-3B-Instruct | |
| language: | |
| - ko | |
| - en | |
| license: cc-by-nc-4.0 | |
| tags: | |
| - speech-to-sft | |
| - emnlp-2026-industry | |
| # expertQA_llama3.2-3b_exp0_medical_seed2 | |
| SFT checkpoint from the EMNLP 2026 Industry Track submission *A Factorial Ablation of a Speech-to-SFT Pipeline: Differential Effects on Data Quality and Downstream Transfer*. | |
| | Field | Value | | |
| |---|---| | |
| | Pipeline condition | **Exp 0** (baseline (no refinement)) | | |
| | Domain | medical | | |
| | Seed | 2 | | |
| | Base model | [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) | | |
| | Training | LoRA (rank 16, α=32, QLoRA 4-bit, lr 2e-4, 3 epochs) | | |
| | Upstream STT | In-house STT (paper main pipeline) | | |
| | License | CC BY-NC 4.0 (research and non-commercial use only) | | |
| **Intended use**: research and non-commercial use only, matching the consent scope of the source audio. | |
| **Companion repository** (code, configs, prompts, sample QA): <https://github.com/flitto/speech-to-sft-ablation-paper> | |