violetxi/qwen35-9b-harvey-v4-notes-conditioned-100m
The violetxi/qwen35-9b-harvey-v4-notes-conditioned-100m is a 9 billion parameter Qwen3.5-based language model developed by violetxi. This model is a specialized checkpoint from an experiment focused on note-conditioned trajectories, trained with a 70%/30% supervised-token mixture. It is optimized for agent generations that involve historical thinking, as evidenced by its evaluation on specific multi-turn tasks.
Loading preview...
Model Overview
The violetxi/qwen35-9b-harvey-v4-notes-conditioned-100m is a 9 billion parameter model built upon the Qwen/Qwen3.5-9B architecture. This specific checkpoint, checkpoint-6134, is the result of an experiment involving 100 million notes and note-conditioned trajectories, trained over two epochs. The training dataset comprised approximately 70 million note labels and 30 million assistant labels from these trajectories, with a total of nearly 200 million tokens exposed after causal shifting.
Key Characteristics & Training
- Base Model: Qwen3.5-9B.
- Training Data: Approximately 100 million supervised tokens, split 70% for note labels and 30% for assistant labels from note-conditioned trajectories.
- Training Objective: Notes use causal next-token prediction, while trajectory loss is applied to assistant labels.
- Context Length: Utilizes a packed sequence length of 16,384 tokens during training.
Evaluation & Performance
This model was evaluated on historical thinking-enabled agent generations, regraded with gpt-5.6-sol using the Harvey per-criterion rubric. The metric measures the fraction of attempts passing every criterion across 250 tasks with 4 samples each.
- 5-Turn Budget: Achieved a 5.00% all-criteria-pass rate on the historical 5-turn evaluation dataset.
- 20-Turn Budget: Achieved an 8.00% all-criteria-pass rate on the historical 20-turn evaluation dataset.
Use Cases
This model is particularly suited for applications requiring:
- Agentic Workflows: Especially those involving multi-turn interactions and historical thinking.
- Note-Conditioned Generation: Scenarios where conditioning on specific notes or trajectories is beneficial.
- Research into Agent Behavior: For exploring and developing agents that can leverage structured 'notes' for improved performance.