laion/nemosci-tasrep-a1mfc-gfistaqc-dev1-scaff-maxeps-swes-r2eg-32b-10pct__Qwen3-32B

Model Overview

This model, laion/nemosci-tasrep-a1mfc-gfistaqc-dev1-scaff-maxeps-swes-r2eg-32b-10pct__Qwen3-32B, is a 32 billion parameter language model derived from the Qwen/Qwen3-32B architecture. It has undergone extensive fine-tuning on a specialized collection of datasets, indicating a focus on particular domains.

Key Fine-tuning Datasets

The model's training involved a diverse set of datasets, suggesting an emphasis on scientific and agent-based tasks:

• Nemotron terminal scientific computing: Likely enhances its ability to process and generate scientific text and code.
• DCAgent experimental traces: Multiple datasets from DCAgent, including exp_tas_repetition_penalty_1.05_traces, a1_multifile_composition, exp-gfi-staqc-embedding-mean-filtered-10K_glm_4.7_traces_jupiter, exp_tas_max_episodes_512_traces, dev_set_part1_10k_glm_4.7_traces_jupiter, and a1_repo_scaffold. These datasets suggest a specialization in understanding and generating agent-like behaviors, complex compositions, and code scaffolding.
• swesmith-sandboxes-with_tests-gpt-5-mini-passed_glm_4.7_traces: Indicates training on code generation and testing scenarios.
• penfever--Kimi-2.5-r2egym_sandboxes-maxeps-32k-10pct: Further reinforces its capabilities in agent environments and potentially reinforcement learning contexts.

Training Details

During training, the model utilized a learning rate of 4e-05, a total batch size of 96 across 96 devices, and was trained for 7 epochs using the AdamW optimizer with a cosine learning rate scheduler. The training was performed using Transformers 4.57.6 and Pytorch 2.9.1+cu130.