| license: apache-2.0 | |
| language: | |
| - en | |
| tags: | |
| - causal-lm | |
| - sft | |
| - logic-reasoning | |
| # TestGeniy accepted logic small-scope SFT | |
| Derived from TestGeniy dialogue SFT v6. Only the last-block attention output and FFN down projection were trained on original train splits from FOLIO, ARC-Challenge, and StrategyQA. Benchmark test/validation rows, synthetic examples, model-generated answers, and MTP were excluded. | |
| | checkpoint | GSM8K | MATH-500 | ARC-Challenge | FOLIO | composite | | |
| |---|---:|---:|---:|---:|---:| | |
| | v6 | 24/100 | 7/100 | 26/100 | 29/100 | 21.5 | | |
| | accepted candidate | 25/100 | 7/100 | 31/100 | 35/100 | 24.5 | | |
| Balanced human preference accuracy on 80 OASST1/HelpSteer2 pairs is unchanged at 55% (OASST1 62.5%, HelpSteer2 47.5%). This release passed the no-regression gate and is separate from production v6. | |