trohrbaugh/BigBang-v1-heretic
trohrbaugh/BigBang-v1-heretic is a 35.1 billion parameter, instruction-tuned causal language model, derived from endless-frontier/BigBang-v1 and decensored using Heretic v1.2.0. It features a 32768 token context length and is optimized for scientific research, reasoning, coding, and tool-use tasks. This model excels at complex problem-solving by leveraging an adversarial, self-evolving synthetic data framework for post-training.
Loading preview...
Model Overview
trohrbaugh/BigBang-v1-heretic is a 35.1 billion parameter language model based on the Qwen 3.6 35B-A3B architecture, fine-tuned by trohrbaugh. This version is a decensored variant of endless-frontier/BigBang-v1, created using the Heretic v1.2.0 tool. The original BigBang-v1 model was developed using an adversarial, self-evolving synthetic data framework, which generates and solves increasingly challenging scientific and technical problems, then evaluates them for correctness and difficulty.
Key Capabilities & Performance
This model demonstrates strong performance across various domains, including scientific research, reasoning, coding, and tool-use. It significantly outperforms its base model and achieves aggregate performance comparable to much larger models like DeepSeek V4 Flash (284B) and DeepSeek V4 Pro (1.6T) on specific benchmarks. Notably, the 'heretic' modification results in a substantial reduction in refusals (0/100 compared to 99/100 for the original model), while maintaining a low KL divergence of 0.1088.
Benchmarks
BigBang-v1-heretic (inheriting BigBang-v1's performance) achieves the highest reported scores among selected 35B models on eight benchmarks, including long-horizon search (BrowseComp, XBench), coding tasks (SWE-Bench Pro, SciCode-V-Sub/Main), scientific research (FS-R, HLE, BioMystery-HS/HD), and AI research (MLE-Bench, PaperBench). It even surpasses DeepSeek V4 Pro Preview (1.6T) on several scientific and AI research benchmarks.
Use Cases
This model is particularly well-suited for applications requiring advanced problem-solving in scientific and technical fields, complex reasoning, code generation, and tool integration. Its decensored nature may also make it suitable for use cases where the original model's refusal rates were prohibitive.