against-the-safety-tiny-aya-rag-patch

A LoRA adapter for Tiny Aya Global (revision af89d219) from the project's third round (candD, "careful patch"). It continues from the main adapter for one epoch (93 steps, 2.2 M tokens) on new episodes that teach it to answer when a single passage holds the answer, to answer both parts of a two-part question, and to treat supplied on-topic passages without the answer as insufficient: 45% new rows (SQuAD 2.0 training-split questions, compound questions, prompt variants) and 55% rehearsal rows from the main adapter's training set. Built for the Lexsi AI Safety and Alignment Hackathon, Track 3, problem statement 4. Team Against-the-safety. Code and plan: https://github.com/Bhaveshmeghwal21/against-the-safety (plan sections 24 to 26).

Status: not the selected model, and not recommended where an unsupported answer is costly. The rules were fixed before training: a candidate could replace the main adapter only if its pooled rate of answers given when no passage in reach holds the answer stayed within 3 points of main's on the development set. This adapter, like candC, did not, so the main adapter stayed the project's model. On fresh questions the same rate was again too high (below). Use it when you need single-passage questions answered and accept more unsupported answers.

Development set (held-out SQuAD training-split articles and HotpotQA dev; ends well)

development set, ends well	main	candC	candD
D1 HotpotQA dev	0.673	0.566	0.629
D2 single-passage	0.325	0.762	0.730
D3 compound	0.510	0.600	0.570
D4 prompt variants	0.497	0.699	0.601
macro average	0.501	0.657	0.633
answers when no passage in reach holds the answer (of 282; lower is better)	17	70	72

Fresh test set (six sets, 739 episodes per system, questions disjoint from every earlier set, opened once)

ends well (episodes)	main	candD (round 3 patch)	candA (round 2)
HotpotQA, fresh (88)	0.602	0.716	0.727
2WikiMultihopQA, fresh (86)	0.651	0.570	0.640
MuSiQue 2-hop, fresh (67)	0.478	0.582	0.627
SQuAD 2.0, fresh questions (166)	0.512	0.645	0.614
RGB, fresh (266)	0.628	0.665	0.658
RGB integration, fresh (66)	0.061	0.152	0.030
macro average of six sets	0.489	0.555	0.549
difference from main, 95% CI		+0.066 [+0.031, +0.103]	+0.061 [+0.028, +0.093]
answers when no passage in reach holds the answer (326 episodes; lower is better)	0.203	0.322	0.233
difference from main, 95% CI		+0.120 [+0.076, +0.163]	+0.031 [-0.003, +0.064]

On fresh SQuAD questions this adapter gave a supported answer to 32 of 57 answerable questions (main: 4, candA: 6) and had 3 invalid turns of 166 (main: 30). It also answered 31 of 109 SQuAD episodes whose passages cannot answer (main: 6, candA: 6). The pre-registered criterion (pooled unsupported answers at most 0.05 above main's) was not met (+0.120).

Use

Loading is the same as for the main adapter (see its card); replace the adapter id with BMEscanor/against-the-safety-tiny-aya-rag-patch. It expects the same system prompt and tool loop (ragharness.protocol.system_prompt, ragharness.rollout).

Limits

The main adapter's limits apply. Two-part answers improved only a little (4 of 27 fresh RGB integration questions answered fully, against 0). When one part of a compound question is missing, the adapter tends to answer the visible part and invent the other. Evaluated sets are small (60 to 120 questions per set); every report gives intervals. Not for safety-critical or legal use.

License

CC BY-NC 4.0, inherited from the base model. Cohere Labs' terms for Tiny Aya apply.