Search for: All records

The BigCode community, an open-scientific collaboration working on the responsible development of Large Language Models for Code (Code LLMs), introduces StarCoder and StarCoderBase: 15.5B parameter models with 8K context length, infilling capabilities and fast large-batch inference enabled by multi-query attention. StarCoderBase is trained on 1 trillion tokens sourced from The Stack, a large collection of permissively licensed GitHub repositories with inspection tools and an opt-out process. We fine-tuned StarCoderBase on 35B Python tokens, resulting in the creation of StarCoder. We perform the most comprehensive evaluation of Code LLMs to date and show that StarCoderBase outperforms every open Code LLM that supports multiple programming languages and matches or outperforms the OpenAI code-cushman-001 model. Furthermore, StarCoder outperforms every model that is fine-tuned on Python, can be prompted to achieve 40% pass@1 on HumanEval, and still retains its performance on other programming languages. We take several important steps towards a safe open-access model release, including an improved PII redaction pipeline and a novel attribution tracing tool, and make the StarCoder models publicly available under a more commercially viable version of the Open Responsible AI Model license. 

Abstract We studied the magnetic excitations in the quasi-one-dimensional (q-1D) ladder subsystem of Sr 14−x Ca x Cu 24 O 41 (SCCO) using Cu L 3 -edge resonant inelastic X-ray scattering (RIXS). By comparing momentum-resolved RIXS spectra with high ( x  = 12.2) and without ( x  = 0) Ca content, we track the evolution of the magnetic excitations from collective two-triplon (2 T) excitations ( x  = 0) to weakly-dispersive gapped modes at an energy of 280 meV ( x  = 12.2). Density matrix renormalization group (DMRG) calculations of the RIXS response in the doped ladders suggest that the flat magnetic dispersion and damped excitation profile observed at x  = 12.2 originates from enhanced hole localization. This interpretation is supported by polarization-dependent RIXS measurements, where we disentangle the spin-conserving Δ S  = 0 scattering from the predominant Δ S  = 1 spin-flip signal in the RIXS spectra. The results show that the low-energy weight in the Δ S  = 0 channel is depleted when Sr is replaced by Ca, consistent with a reduced carrier mobility. Our results demonstrate that off-ladder impurities can affect both the low-energy magnetic excitations and superconducting correlations in the CuO 4 plaquettes. Finally, our study characterizes the magnetic and charge fluctuations in the phase from which superconductivity emerges in SCCO at elevated pressures.