Seerror / Research
Publications

Research Papers

Open-access technical papers from Seerror Technologies. Indexed for readers and for scholarly crawlers, with a public PDF, abstract, and live view and download counts.

4 Papers live
Zenodo DOI
Scholar Ready
Library
Preprint · 28 Oct 2026 · Official record: Zenodo · v407 · current

Jay Tiwari, Seerror Technologies, Bengaluru. Third model, corpus v407, 140,813 conversations, 1.11 GB GGUF. On the previous 100-prompt suite, 86 of 100 answers are safe (95% CI 78–92), against 87 for v349 and 23 for the base model. On the original 70 prompts the three models are safe on 58, 59 and 57. Ratings were made by an AI assistant; the author did not re-rate this version. Hand-written test: 76% of 252 answers safe (83% English, 81% Hinglish, 52% Hindi). The model still invents law sections.

Views — Downloads —
Read paper → PDF Zenodo DOI
Preprint · 9 Oct 2026 · Official record: Zenodo · v349 · previous

Jay Tiwari, Seerror Technologies, Bengaluru. QLoRA fine-tune of Qwen3-1.7B on 107,575 conversations; 1.11 GB GGUF. Manual audit of 100 prompts: 87 safe (95% CI 79–92) against 35 for the base model. On the 70 prompts shared with v224 it is 59 safe against 58, and it still fabricates legal sections. Held-out automatic score looks near-perfect, but 181 answers copy training text. Cohen’s κ = 0.12.

Views — Downloads —
Read paper → PDF Zenodo DOI
Preprint · 2 Oct 2026 · Official record: Zenodo · v224 · previous

Jay Tiwari, Seerror Technologies, Bengaluru. Fine-tuned Qwen3-1.7B with QLoRA on 36,669 records; deployed as a 1.11 GB GGUF model. Manual audit of 70 prompts: 58 safe (82.9%), 12 wrong. Automatic harness 90%, Cohen's κ = 0.10. Errors concentrate in fabricated legal sections and Hinglish prompts.

Views — Downloads —
Read paper → PDF Zenodo DOI

Interested in collaborating on research or responsible disclosure? Reach out with a short note about your scope and intent.

Contact the team