Model is currently being retrained from scratch, at the base model stock model.

✨ Model Summary

A lightweight, reasoning-focused language model fine-tuned on a curated blend of instruction data and character-driven prompts. Designed for both single-turn logic and multi-turn, in-character reasoning tasks.

Reasoning uses: <think>\n ... \n</think>\n response here

🧠 Training Dataset

The model was fine-tuned on:

  • 8,090 entries of mixed single-turn and multi-turn reasoning instructions, including problem-solving, logical inference, and step-by-step tasks.

  • 3,178 entries of mixed single-turn and multi-turn in-character reasoning / RP-style instruction.

  • 2,227 Multi-turn entries derived from character card reasoning.

Downloads last month
12
Safetensors
Model size
12.2B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Nitral-Archive/Irix-12B-Reasoner-Base

Finetuned
(1)
this model
Quantizations
3 models