Lex Fridman Interviewer — Nemotron-3-Nano-4B GRPO v12 (merged weights)
Full merged weights of Nemotron-3-Nano-4B fine-tuned to ask Lex Fridman-style interview questions.
Training
- LoRA v1 (r=64, LR=2e-4, 1 epoch, 4,772 pairs) → score 0.733
- GRPO v12 (reward_v12, 200 steps, LR=5e-6) → score 0.760
Eval (functional judge: on_topic × uses_guest × probing)
| Model | Score | uses_guest | probing |
|---|---|---|---|
| Base | 0.653 | 48% | 84% |
| LoRA v1 | 0.733 | 56% | 92% |
| This model | 0.760 | 60% | 96% |
Note on ONNX
NemotronH (Mamba-2 hybrid) cannot be exported to ONNX — the 38 SSM layers
use compiled CUDA kernels with no ONNX equivalent.
Use this model with vllm or llama.cpp.
Reward Design (reward_v12)
uses_guest_logit^0.67 × probing_logit^0.33 + lexical_bonus
Continuous judge logits from Qwen3.5-4B.
- Downloads last month
- 1,258