Van Nyaya GGUF (Tiny Aya Global + Van Nyaya SFT LoRA, merged)

CPU-ready GGUF of the Van Nyaya headline model: a grounded English + Hindi assistant for Indian forest and environment law (332 sections from the Indian Forest Act 1927, Wildlife (Protection) Act 1972, Forest (Conservation) Act 1980 / Van Adhiniyam 2023, Forest Rights Act 2006, Environment (Protection) Act 1986).

Held-out results (n = 164, sections never seen in training):

correct refusal on unanswerable questions: 10% base -> 100%
English token F1: 0.294 -> 0.359 (+22%)
Hindi token F1: 0.265 -> 0.306 (+15%)
false refusals on answerable questions: 0 -> 3 of 124

Run it:

llama-cli -m model.gguf -sys "You are Van Nyaya, an assistant on Indian forest and environment law. Answer only from the section given, cite it, and say so if the section does not answer the question." -p "Section: <paste section text>\n\nQuestion: <your question>"

Adapter: https://huggingface.co/AkshayWagh/van-nyaya-tiny-aya-sft-lora Code: https://github.com/akshay-ira-2026/van-nyaya Write-up: https://huggingface.co/spaces/AkshayWagh/van-nyaya

Not legal advice.

Evaluation interpretation and toolkit attribution

The refusal figure is a lexical refusal-marker proxy on 40 mismatched-context questions, not human-verified legal correctness or a semantic safety guarantee. Token F1 measures reference overlap, not legal validity. CuratorKIT was run for curation; custom LoRA scripts perform training (AlignTune itself was not executed). Actual AuditKIT F1Score and Abstention offline cross-checks are available in the evaluation dataset and the reproducible script. AuditKIT native normalization differs from the headline custom scorer. See the full write-up for limitations.