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.jsonppo_parameter_change_proof.jsonppo_execution_proof.jsonppo_validation_predictions.csv
Serving
TensorTalk production serves this model through RunPod Serverless vLLM with an OpenAI-compatible endpoint.
Expected served model name:
