violetxi/qwen35-9b-harvey-v4-notes-conditioned-30m
The violetxi/qwen35-9b-harvey-v4-notes-conditioned-30m is a 9 billion parameter Qwen3.5-based language model developed by violetxi, fine-tuned specifically for note-conditioned trajectory experiments. This model was trained with a unique supervised token mixture focusing on note labels and assistant labels from note-conditioned trajectories. It is optimized for agentic reasoning tasks, demonstrating specific performance on historical thinking-enabled agent generations with a 32768 token context length.
Loading preview...
Model Overview
The violetxi/qwen35-9b-harvey-v4-notes-conditioned-30m is a 9 billion parameter model built upon the Qwen/Qwen3.5-9B architecture. It represents a specific checkpoint from an experiment focusing on 30 million notes + note-conditioned trajectory training, utilizing a 70% note label and 30% assistant label supervised token mixture without KL regularization.
Training Details
This model was trained for 2 epochs, completing 1,836 optimizer updates. The training dataset included approximately 30 million supervised tokens per epoch, with notes using causal next-token prediction and trajectory loss applied to assistant labels. It used a packed sequence length of 16,384 and a learning rate of 5e-06.
Evaluation and Performance
The model's performance was evaluated on historical thinking-enabled agent generations, regraded with gpt-5.6-sol using the Harvey per-criterion rubric. It achieved an all-criteria-pass rate of 5.50% with a 5-turn budget and 7.00% with a 20-turn budget on specific evaluation datasets. These evaluations used the historical 5/20-turn glob/grep/read protocol. Further closed-book recall probes were conducted, involving 7,933 probes with deterministic scoring.
Key Characteristics
- Base Model: Qwen3.5-9B.
- Specialized Training: Fine-tuned on 30 million notes and note-conditioned trajectories.
- Agentic Reasoning: Optimized for tasks requiring historical thinking and agentic generation.
- Context Length: Supports a context length of 32,768 tokens.