The DOI matches its Crossref record; 53 of the 78 references it cites were found in a registry or an index; 26 of the 38 sources its claims cite were checked for support.
Training Compute-Optimal Large Language Models
Open accessAbstract
We investigate the optimal model size and number of tokens for training a transformer language model under a given compute budget. We find that current large language models are significantly undertrained, a consequence of the recent focus on scaling language models whilst keeping the amount of training data constant. By training over 400 language models ranging from 70 million to over 16 billion parameters on 5 to 500 billion tokens, we find that for compute-optimal training, the model size and the number of training tokens should be scaled equally: for every doubling of model size the number of training tokens should also be doubled. We test this hypothesis by training a predicted compute-optimal model, Chinchilla, that uses the same compute budget as Gopher but with 70B parameters and 4$\times$ more more data. Chinchilla uniformly and significantly outperforms Gopher (280B), GPT-3 (175B), Jurassic-1 (178B), and Megatron-Turing NLG (530B) on a large range of downstream evaluation tasks. This also means that Chinchilla uses substantially less compute for fine-tuning and inference, greatly facilitating downstream usage. As a highlight, Chinchilla reaches a state-of-the-art average accuracy of 67.5% on the MMLU benchmark, greater than a 7% improvement over Gopher.
Overview
CoChat Summary
TL;DR
The paper estimates compute-optimal scaling for transformer language models and finds that model size and training tokens should be increased in roughly equal proportions as compute grows. It validates this by training Chinchilla, a 70B model on 1.4T tokens, which outperforms much larger models such as Gopher (280B) under the same compute budget.
Key findings
- Across three fitting approaches, the scaling exponents for optimal parameters \(N_{opt} \propto C^a\) and tokens \(D_{opt} \propto C^b\) are near equal: Approach 1 gives \(a=0.50, b=0.50\); Approach 2 gives \(a=0.49, b=0.51\); Approach 3 gives \(a=0.46, b=0.54\), in contrast to Kaplan et al. (2020), which reported \(a=0.73, b=0.27\).
- Chinchilla (70B parameters, 1.4T tokens) uses the same compute budget as Gopher (280B, 300B tokens) and outperforms Gopher, GPT-3 (175B), Jurassic-1 (178B), and Megatron-Turing NLG (530B) on a large range of downstream tasks.
- Chinchilla reaches 67.6% average 5-shot accuracy on MMLU, improving on Gopher by 7.6%; it also improves BIG-bench average accuracy to 65.1% versus 54.4% for Gopher, and achieves 77.4% on LAMBADA zero-shot versus 74.5% for Gopher.
- The paper concludes that current large language models are significantly undertrained and that a 175B model should be trained on over 4.2 trillion tokens, while a 280B Gopher-like model should be trained on 6.8 trillion tokens.
Methods
- The authors train over 400 language models ranging from 70 million to over 16 billion parameters on 5 to 500 billion tokens, then fit scaling relationships using three approaches: fixing model sizes and varying training tokens, IsoFLOP profiles, and fitting a parametric loss function.
- They use the resulting compute-optimal frontier to train Chinchilla, a 70B parameter autoregressive transformer, on 1.4 trillion tokens using the same compute budget as Gopher, with architecture and training setup based on Gopher but with changes including AdamW, a modified SentencePiece tokenizer without NFKC normalization, and a different MassiveText subset distribution.
- Evaluation covers language modelling, reading comprehension, question answering, common sense, MMLU, BIG-bench, gender bias, and toxicity, largely following the evaluation setup of Rae et al. (2021).
Limitations
- The paper states that due to the cost of training large models, it has only two comparable large-scale training runs (Chinchilla and Gopher) and no additional tests at intermediate scales.
- It assumes the efficient computational frontier follows a power-law relationship, but observes concavity in \(\log N_{opt}\) at high compute budgets, suggesting it may still overestimate the optimal size of large models.
- All training runs for the analysis used less than one epoch of data, so future work may consider the multiple-epoch regime; the paper also notes that language modelling comparisons may be affected by train/test leakage because Chinchilla is trained on 4× more data than Gopher, and its bias/toxicity evaluations are not comprehensive.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
22 Valid · 5 Not valid · 9 Not checked · 88 Own results and interpretations
Every source the paper cites for this claim was checked and supports it.
In the paper
“In Section D.4, we show a head-to-head comparison at 10 21 FLOPs, using the model size recommended by our analysis and by the analysis of Kaplan et al. (2020)-using the model size we predict has a clear advantage.”
What the sources say
- Supportedcites Kaplan et al. (2020)
Section 6.1 (p. 15) gives Kaplan et al.’s recommended model size as a function of compute budget, so their analysis can supply a model size for a comparison at 10^21 FLOPs. The claimed advantage of the citing authors’ model size is presented as the citing authors’ own finding, not as a finding attributed to Kaplan et al. (LLM, conf=0.90)
boostedstates a causeSingle case- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources that could be checked couldn’t confirm or dispute this claim.
In the paper
“the third term should scale as the convergence rate of these methods, which is lower-bounded by 1 𝐷 1 2 (Robbins and Monro, 1951) (and may attain the bound).”
What the sources say
- Could not verifycites Robbins and Monro, 1951 sourceExperimental — accuracy not yet measured
The Summary (p. 1) and Convergence theorems (pp. 2–5) establish convergence of a stochastic approximation procedure, but the readable passages do not give a 1/√D lower bound for the convergence rate of stochastic first-order methods. I searched for the rate, lower bound, and efficiency; several formulas in our copy are undecoded, so the claimed bound cannot be verified from it. (LLM, conf=0.72)
hedgedstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Could not verify
Every source the paper cites for this claim was checked and supports it.
In the paper
“On the filtered version, Chinchilla lags behind the open book SOTA (Izacard and Grave, 2020) by only 7.9%.”
What the sources say
- Supportedcites Izacard and Grave, 2020 source
Table 2 (4.2 Technical Details, p. 7) reports a 72.5 TriviaQA test exact-match score for the paper’s open-book retriever–reader system, and 4.3 Results (p. 7) describes it as outperforming the prior state of the art. The citation supports the open-book SOTA benchmark it identifies; the 7.9-point comparison with Chinchilla is the citing authors’ calculation, not a result reported in this source. (LLM, conf=0.86)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“Our approach leads to considerably different results than that of Kaplan et al. (2020).”
What the sources say
- Supportedcites Kaplan et al. (2020)
The citation identifies Kaplan et al.’s approach and results, rather than asking that paper to assess the citing authors’ approach. Section 3 (p. 7) describes the models Kaplan et al. trained, and the Discussion (p. 18) states their scaling results. The comparison with the citing authors’ results is their own conclusion. (LLM, conf=0.88)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“From Kaplan et al. (2020), we find that the optimal model size should be 4.68 billion parameters.”
What the sources say
- Partly supportedcites Kaplan et al. (2020)Experimental — accuracy not yet measured
Section 6.1 (p. 15) says that each compute budget has an associated optimal model size and that the authors fit this relationship with a power law. Our copy does not give a readable coefficient or a value at 10^21 FLOPs, so it does not verify the specific figure of 4.68 billion parameters. (LLM, conf=0.83)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
45 more claims · 2 with their cited sources checked
Create a free account to see all 124 claims and to verify the 2 references still open.
- It may state two separate claims as one. 48
- Probably not a substantial claim. 34
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Initially, scaling efforts focused on training-time expansion through increases in model and dataset size [14,18].”
From A Survey of Large Language Models (2026) · cites this paper as [18]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“In SIMA, however, we are drawing inspiration from the lesson of large language models that training on a broad distribution of data is the most effective way to make progress in general AI (e.g., Brown et al., 2020;Hoffmann et al., 2022;OpenAI, 2023;Anil et al., 2023;Gemini Team et al., 2023).”
From Scaling Instructable Agents Across Many Simulated Worlds (2024) · cites this paper as Hoffmann et al., 2022
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“For example, Touvron et al. (2023a) are forced to pull numbers from the PaLM and Chinchilla technical reports (Chowdhery et al., 2023; Hoffmann et al., 2022) directly for baseline comparison, with no option to rerun these models on Llama's evaluation implementations since PaLM and Chinchilla are not publicly available.”
From Lessons from the Trenches on Reproducible Evaluation of Language Models (2024) · cites this paper as Hoffmann et al., 2022
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Transformers are widely known to under-perform with insufficient training compute compared to RNNs [53, 40].”
From End-to-End Test-Time Training for Long Context (2025) · cites this paper as [40]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Chinchilla outperformed the much larger Gopher [Rae et al., 2021], GPT-3 [Brown et al., 2020], and others [Smith et al., 2022] across a broad benchmark suite”
From Training Compute-Optimal Transformer Models: A Review of Scaling Laws, Allocation, and Practice (2026) · cites this paper as Hoffmann et al., 2022
Experimental — accuracy not yet measured
22 more claims about this paper
Create a free account to see every claim other papers make about this one.
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Training Compute-Optimal Large Language Models · p. 1
“Chinchilla reaches a state-of-the-art average accuracy of 67.5% on the MMLU benchmark”
§ 4.2.1. Language modelling · p. 11
“| Chinchilla 5-shot | 67.6% |”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 4.2.5. Common sense · p. 12
“We find that Chinchilla outperforms both Gopher and GPT-3 on all tasks”
§ 4.2.7. Gender bias and toxicity · p. 13
“| PIQA | 81.8% | 81.8% | 81.0% | 82.0% | 90.1% |”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ D.3. Predicted compute optimal frontier for all three methods · p. 26
“| 175 Billion | 4.54e+24 | 4.3 Trillion | 1.26e+24 | 12.0 Trillion |”
§ D.3. Predicted compute optimal frontier for all three methods · p. 26
“| 67 Billion | 6.88e+23 | 1.7 Trillion | 1.71e+24 | 4.1 Trillion |”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 3.4. Optimal model scaling · p. 8
“Unless one has a compute budget of 10 26 FLOPs (over 250 the compute used to train Gopher )”
§ 3.4. Optimal model scaling · p. 8
“| 1 Trillion | 1.27e+26 | 221 3 | 21.2 Trillion |”
Citations
Cites
This paper’s reference list could not be split into single entries.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017) · arXiv · doi:10.48550/arxiv.1706.03762 · Cited by 120,404 · cited 1×
“These large autoregressive transformers (Vaswani et al., 2017) have demonstrated impressive performance on many tasks using a variety of evaluation protocols such as zero-shot, few-shot, and fine-tuning.”
Adam: A Method for Stochastic Optimization
Kingma, D. P., & Ba, J. (2014) · arXiv · doi:10.48550/arxiv.1412.6980 · Cited by 98,946 · cited 6×
“We find that, independent of the learning rate schedule, AdamW trained models outperform models trained with Adam.”
InfluentialLanguage Models are Few-Shot Learners
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. (2020) · arXiv · doi:10.48550/arxiv.2005.14165 · Cited by 23,372 · cited 9×
“We find that Chinchilla outperforms both Gopher and GPT-3 on all tasks and outperforms MT-NLG 530B on all but one task—see Table 8.”
InfluentialExploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer
Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., & Liu, P. J. (2019) · arXiv · doi:10.48550/arxiv.1910.10683 · Cited by 10,929 · cited 1×
“In Appendix C, we reproduce the IsoFLOP analysis on two additional datasets: C4 [40] and GitHub code [38].”
A Stochastic Approximation Method
Robbins, H., & Monro, S. (1951) · The Annals of Mathematical Statistics · doi:10.1214/aoms/1177729586 · Cited by 9,643 · cited 1×
“Finally, given that it corresponds to early stopping in stochastic first order methods, the third term should scale as the convergence rate of these methods, which is lower-bounded by 1 𝐷 1 / 2 (Robbins and Monro, 1951) (and may attain the bound).”
Scaling Laws for Neural Language Models
Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., & Amodei, D. (2020) · arXiv · doi:10.48550/arxiv.2001.08361 · Cited by 9,550 · cited 17×
“Our approach leads to considerably different results than that of Kaplan et al. (2020).”
InfluentialRobust estimation of a location parameter
Huber, P. J. (1964) · The Annals of Mathematical Statistics · doi:10.1214/aoms/1177703732 · Cited by 7,228
On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜
Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021) · Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency · doi:10.1145/3442188.3445922 · Cited by 6,988 · cited 1×
“Large Language Models carry potential risks such as outputting offensive language, propagating social biases, and leaking private information (Bender et al., 2021; Weidinger et al., 2021).”
In-Datacenter Performance Analysis of a Tensor Processing Unit
Jouppi, N. P., Young, C., Patil, N., Patterson, D., Agrawal, G., Bajwa, R., Bates, S., Bhatia, S., Boden, N., Borchers, A., et al. (2017) · Proceedings of the 44th Annual International Symposium on Computer Architecture · doi:10.1145/3079856.3080246 · Cited by 4,537 · cited 1×
“All models in this analysis have been trained on TPUv3/TPUv4 (Jouppi et al., 2017) with JAX (Bradbury et al., 2018) and Haiku (Hennigan et al., 2020).”
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W.-T., Rocktäschel, T., et al. (2020) · arXiv · doi:10.48550/arxiv.2005.11401 · Cited by 3,101
Cited by
Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A. C., Lo, W.-Y., et al. (2023) · 2023 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv51070.2023.00371 · Cited by 11,953
“Empirical trends show this behavior improving with model scale, dataset size, and total training compute [56, 10, 21, 51].”
BackgroundLLaMA: Open and Efficient Foundation Language Models
Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.-A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al. (2023) · arXiv · doi:10.48550/arxiv.2302.13971 · Cited by 4,602
Large Language Models Encode Clinical Knowledge
Singhal, K., Azizi, S., Tu, T., Mahdavi, S. S., Wei, J., Chung, H. W., Scales, N., Tanwani, A., Cole-Lewis, H., Pfohl, S., et al. (2022) · arXiv · doi:10.48550/arxiv.2212.13138 · Cited by 4,083
“Large language models (LLMs) Over the past few years, LLMs have shown impressive performance on natural language processing (NLP) tasks [12, 14, 15, 30, 69, 70, 73, 89, 91, 99].”
BackgroundLarge language models encode clinical knowledge
Singhal, K., Azizi, S., Tu, T., Mahdavi, S. S., Wei, J., Chung, H. W., Scales, N., Tanwani, A., Cole-Lewis, H., Pfohl, S., et al. (2023) · Nature · doi:10.1038/s41586-023-06291-2 · Cited by 4,083
“Large language models (LLMs) Over the past few years, LLMs have shown impressive performance on natural language processing (NLP) tasks [12, 14, 15, 30, 69, 70, 73, 89, 91, 99].”
BackgroundPaLM: Scaling Language Modeling with Pathways
Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H. W., Sutton, C., Gehrmann, S., et al. (2022) · arXiv · doi:10.48550/arxiv.2204.02311 · Cited by 2,590
“, 2021), Chinchilla (Hoffmann et al., 2022), Megatron–Turing NLG (Smith et al.”
BackgroundMethodResultInfluential
Metadata
- Authors
- Hoffmann, J., Borgeaud, S., Mensch, A., Buchatskaya, E., Cai, T., Rutherford, E., Casas, D. D. L., Hendricks, L. A., Welbl, J., Clark, A., Hennigan, T., Noland, E., Millican, K., Driessche, G. V. D., Damoc, B., Guy, A., Osindero, S., Simonyan, K., Elsen, E., Vinyals, O., Rae, J., & Sifre, L.
- Venue
- Advances in Neural Information Processing Systems 35
- Published
- 2022
- Pages
- 30016-30030
- ISSN
- 1049-5258 (ISSN-L)
- OpenAlex
- W7133225605
- Cited by
- 932
- References
- 78
- Citations by year
- 2026: 242 · 2025: 188 · 2024: 248 · 2023: 226 · 2022: 26
- FWCI
- 96.42
- Type
- conference paper
- Language
- en
- Open access
- green
- arxiv.org
- Locations
- neural information processing systems (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University)
- Funders
- DeepMind
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Keywords
- large language models, computational efficiency, training efficiency, model compression, inference efficiency
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-27, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-27 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Training Compute-Optimal Large Language Models. https://app.cochat.ai/papers/10.52202/068431-2176
- The paper
- Hoffmann, J., Borgeaud, S., Mensch, A., Buchatskaya, E., Cai, T., Rutherford, E., Casas, D. D. L., Hendricks, L. A., Welbl, J., Clark, A., Hennigan, T., Noland, E., Millican, K., Driessche, G. V. D., Damoc, B., Guy, A., Osindero, S., Simonyan, K., Elsen, E., . . . Sifre, L. (2022). Training Compute-Optimal Large Language Models. Advances in Neural Information Processing Systems 35. https://doi.org/10.52202/068431-2176
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.