promotion/NBPO-Qwen2.5-7B-WildChecklists-2ep

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 16, 2026Architecture:Transformer Featherless Exclusive Cold

NBPO-Qwen2.5-7B-WildChecklists-2ep is a 7.6 billion parameter language model fine-tuned from Qwen/Qwen2.5-7B-Instruct. It utilizes a Nash bargaining over game-value improvements approach with an adaptive self-play reference, where dual multipliers are fitted per prompt. This model is specifically trained on the WildChecklists dataset, where each prompt includes its own checklist items for granular judging. It is a research checkpoint designed to explore different aggregation rules in training, offering a unique perspective on fine-tuning methodologies.

Loading preview...

Overview

NBPO-Qwen2.5-7B-WildChecklists-2ep is a 7.6 billion parameter research checkpoint model, fine-tuned from Qwen/Qwen2.5-7B-Instruct. This model explores a novel training approach called Nash Bargaining over game-value improvements (NBPO), utilizing an adaptive self-play reference and fitting dual multipliers separately for each prompt. It is part of a comparative study of three models where the only difference is the aggregation rule used during training.

Key Capabilities and Training

This model was retrained using the "Back to Blackwell" (PROSPER) protocol on the WildChecklists dataset. This dataset is unique because every prompt comes with its own checklist items, and a judge scores one item at a time. Key training details include:

  • Prompts: 522 unique prompts.
  • Learner Pairs: 14,616 pairs from 8 sampled responses per prompt.
  • Pipeline Epochs: 2, with 228 optimizer steps.
  • Max Sequence Length: 2048 tokens.
  • Judge: A local open-weight judge was used for training signal and evaluation.

Evaluation and Limitations

The model was evaluated using greedy generation against benchmark baseline answers. On Arena-Hard, it scored 0.6333, and on AlpacaEval, it scored 0.4124. It's important to note two key limitations regarding these scores:

  1. Not Leaderboard Scores: The evaluations use a local open-weight judge, not the gpt-4-1106 or weighted_alpaca_eval_gpt4_turbo typically used for official leaderboards. Therefore, absolute values are not directly comparable to published leaderboard numbers, though within-judge comparisons between the three arms are valid.
  2. Statistical Significance: Differences between this NBPO arm and the PROSPER or Fixed-reference Nash arms are not statistically resolved, meaning performance variations are within the margin of error for most benchmarks. General-capability benchmarks like MMLU and HellaSwag show no significant separation from the base model or other arms.

Good for

  • Researchers interested in advanced fine-tuning methodologies, particularly Nash bargaining and adaptive reference training.
  • Experiments with models trained on granular, checklist-based evaluation datasets.
  • Comparative studies of different aggregation rules in reinforcement learning from human feedback (RLHF) or similar training paradigms.

Finetunes
1 models