File size: 1,022 Bytes
a0a381e
 
 
 
 
 
 
 
 
 
 
0a4a501
a0a381e
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
# Representative source code for VR and VRP training

This folder provides representative source code for Qwen3-0.6B noSFT training under PPO and GRPO with VR and VRP rewards.

The released notebooks correspond to the reward construction described in the paper:

- VR uses the base verifier reward.
- VRP extends VR with hard gate checking and soft penalty shaping.
- MultiEvalSumViet2 outputs scores in the order Faithfulness, Coherence, and Relevance.
- The scalar reward uses 0.5F + 0.3R + 0.2C.
- The VRP hard gate uses only Faithfulness, Coherence, Relevance, and length score.


Files:

- `ppo_qwen3_06b_noSFT_VR_fulltrain.ipynb`: PPO with VR.
- `ppo_qwen3_06b_noSFT_VRP_fulltrain.ipynb`: PPO with VRP.
- `grpo_qwen3_06b_noSFT_VR_fulltrain.ipynb`: GRPO with VR.
- `grpo_qwen3_06b_noSFT_VRP_fulltrain.ipynb`: GRPO with VRP.

These scripts use the full training file after removing examples with missing source articles. Other model families use the same reward construction with model specific configuration changes.