PJMixers-Dev/LLaMa-3.2-Instruct-JankMix-v0.2-SFT-3B was further trained using KTO (with apo_zero_unpaired loss type) using a mix of instruct, RP, and storygen datasets. I created rejected samples by using the SFT with bad settings (including logit bias) for every model turn.
The model was only trained at max_length=6144, and is nowhere near a full epoch as it eventually crashed. So think of this like a test of a test.
W&B Training Logs
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 21.69 |
| IFEval (0-Shot) | 65.04 |
| BBH (3-Shot) | 22.29 |
| MATH Lvl 5 (4-Shot) | 11.78 |
| GPQA (0-shot) | 2.91 |
| MuSR (0-shot) | 4.69 |
| MMLU-PRO (5-shot) | 23.42 |