---
license: apache-2.0
base_model: nvidia/Nemotron-3-4B-Base
library_name: transformers
pipeline_tag: text-generation
tags:
  - grpo
  - interview
  - lex-fridman
  - nemotron
  - mamba
---

# 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
1. **LoRA v1** (r=64, LR=2e-4, 1 epoch, 4,772 pairs) → score 0.733  
2. **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.