nfdlh/tensortalk-v2

TensorTalk v2

TensorTalk v2 is the Stage 3 PPO fine-tuned Qwen3-8B model for the UM FSKTM student handbook assistant.

This repository contains the merged/full inference model used by RunPod vLLM. It intentionally excludes adapter-only files from the upload so the serving endpoint can load model.safetensors directly.

Training Summary

• Base model: Qwen3-8B
• Fine-tuning stage: Stage 3 PPO
• Reward type: balanced rule-based preference reward function
• Training rows: 900
• Validation rows: 100
• PPO epochs: 2
• Training log records: 900
• Trainable ratio during PPO: about 0.92%

The included proof files record the training run, parameter-change audit, and validation predictions:

• ppo_training_manifest.json
• ppo_parameter_change_proof.json
• ppo_execution_proof.json
• ppo_validation_predictions.csv

Serving

TensorTalk production serves this model through RunPod Serverless vLLM with an OpenAI-compatible endpoint.

Expected served model name:

nfdlh/tensortalk-v2