Datasets:
|
Download README.md from SushiOrbit/tinyaya-sft: direct link, hf CLI and curl.
- Browser
- Download file 3.24 kB
-
https://huggingface.co/datasets/SushiOrbit/tinyaya-sft/resolve/main/README.md
- Command line
-
hf download hf://datasets/SushiOrbit/tinyaya-sft/README.md
-
curl -L -o README.md https://huggingface.co/datasets/SushiOrbit/tinyaya-sft/resolve/main/README.md
3.24 kB
| license: cc-by-nc-4.0 | |
| language: | |
| - en | |
| task_categories: | |
| - text-generation | |
| tags: | |
| - sft | |
| - medical | |
| - curatorkit | |
| - lexsi-hackathon | |
| pretty_name: Tiny Aya medical and control fine-tuning sets | |
| size_categories: | |
| - 1K<n<10K | |
| configs: | |
| - config_name: sft_medical | |
| data_files: data/sft_medical/train.jsonl | |
| - config_name: sft_control | |
| data_files: data/sft_control/train.jsonl | |
| # Tiny Aya fine-tuning sets: medical (MedMCQA explanations) and token-matched control (Alpaca-cleaned) | |
| **Lexsi AI Safety and Alignment Hackathon, Track 2.** Team: Sushovit Nanda and Manish Shaw. Project: *Safe from Fine-Tune to Phone*. Claim: Benign medical LoRA doubles Tiny Aya Global's attack success (18.2 % to 36.3 %); a run-time steering vector cuts it to 26.3 % and survives GGUF quantization. | |
| All repositories of this project are public and un-gated. No model completion on a harmful prompt is published anywhere; prompt sets carry prompts and labels only, and evaluation logs are text-free records. Write-up, code and reproduction instructions: https://github.com/SushovitNanda/Safe-from-Fine-Tune-to-Phone. Collections: [models](https://huggingface.co/collections/SushiOrbit/safe-from-fine-tune-to-phone-models), [datasets](https://huggingface.co/collections/SushiOrbit/safe-from-fine-tune-to-phone-datasets). | |
| **This repository** holds the two curated supervised fine-tuning sets used to induce and control for safety drift: 4,815 medical rows built from MedMCQA questions with explanations, and 5,004 general-instruction rows from Alpaca-cleaned, trimmed (seed 43) so that the two token totals agree within 1 % (2.650 M against 2.676 M templated tokens). Both passed CuratorKIT 0.1.21: MinHash deduplication (3-gram shingles, 128 permutations, Jaccard at or above 0.85), a Detoxify toxicity gate at 0.5, and a 10 to 4,096-token length rule; the medical set went from 4,998 rows to 4,843 before trimming (139 rejected by the toxicity gate for anatomy, abuse and death, about 16 duplicates), the control set lost 50 rows. `rejected.jsonl` lists the rows the gate removed; `stats.json` and `lexsi_provenance.json` are CuratorKIT's own reports; `sft_sharegpt.jsonl` is the ShareGPT export. A three-way leak check (exact match, MinHash, embedding cosine at 0.92) against every evaluation set was run before training. | |
| Sources: MedMCQA (openlifescienceai/medmcqa, Apache-2.0) and Alpaca-cleaned (yahma/alpaca-cleaned, CC-BY-NC-4.0). This derived release is CC-BY-NC-4.0. | |
| Fine-tuning recipe that used these sets: LoRA rank 16, alpha 32, dropout 0.05 on all linear modules; one epoch; learning rate 2e-4, cosine schedule, 3 % warm-up; batch 1 with 16 accumulation steps; FP16 mixed precision; seed 13; loss on assistant tokens only. The resulting adapters are in the model collection of this project. | |
| ```python | |
| from datasets import load_dataset | |
| ds = load_dataset("SushiOrbit/tinyaya-sft", "sft_medical") | |
| ``` | |
| `rejected.jsonl` holds source rows (MedMCQA explanations, Alpaca instructions) that the Detoxify gate removed for anatomy, abuse or death vocabulary; they are public source data, not model output. CuratorKIT's own per-set statistics and provenance records are in `data/<set>/stats.json`, `manifest.json` and `lexsi_provenance.json`. | |