Ordis-1.5B-V355-VarGH: A Practical 1.5B Model
Ordis-1.5B-V355-VarGH, developed by sugiken, is a 1.5 billion parameter model fine-tuned from Qwen2.5-1.5B-Instruct. Its core focus is on enhancing practical capabilities like anti-hallucination, honest refusal ("I don't know"), and structured reasoning, rather than maximizing standard benchmark scores. The model was trained using LoRA with a 4-stage Progressive Identity Training (PIT) pipeline, involving over 16 controlled experiments.
While standard benchmarks show a slight "alignment tax" (minor regression compared to the base model), Ordis excels in areas not typically measured. Notably, it achieved +1.02% on TruthfulQA MC2, demonstrating improved truthfulness. It also shows strong performance in CLadder Causal Reasoning, scoring 54.33% overall, comparable to larger models like LLaMA-6.7B and GPT-3.5. The model exhibits a significant shift in 0-shot counterfactual reasoning (CRASS AI), becoming more conservative due to its anti-hallucination training.
Key Capabilities
- Anti-Hallucination: Uncertainty derived from reasoning, not just safety templates.
- Three-Layer Cognition: Ability to assess knowledge, acknowledge unknowns, and verify before responding.
- Structured Self-Correction: Capable of acknowledging, attributing, correcting, and verifying information.
- Causal Reasoning: Demonstrates cross-domain causal structure transfer.
Good For
- Applications requiring high reliability and truthfulness.
- Scenarios where honest refusal and structured reasoning are critical.
- Tasks benefiting from causal reasoning and self-correction.
- Use cases where a smaller, more cognitively robust model is preferred over raw performance on generic benchmarks.