AI Policy Wiki
Dashboard

LLaMA: Open and Efficient Foundation Language Models (Touvron et al., Meta AI, February 2023)

high confidence · updated 2026-08-01

Meta AI's paper introducing the LLaMA family (7B–65B), trained only on publicly available data, reporting that LLaMA-13B outperforms GPT-3 (175B) on most benchmarks and LLaMA-65B is competitive with Chinchilla-70B and PaLM-540B. Argues that for a fixed inference budget the preferred model is the smaller one trained on more tokens.

"LLaMA: Open and Efficient Foundation Language Models" is a paper submitted to arXiv on February 27, 2023 (arXiv:2302.13971) by fourteen authors at Meta AI, with Hugo Touvron, Thibaut Lavril, Gautier Izacard, Edouard Grave and Guillaume Lample marked as equal contributors. It introduces a family of foundation language models ranging from 7B to 65B parameters and states two claims in its abstract: "that it is possible to train state-of-the-art models using publicly available datasets exclusively, without resorting to proprietary and inaccessible datasets," and that "LLaMA-13B outperforms GPT-3 (175B) on most benchmarks, and LLaMA-65B is competitive with the best models, Chinchilla-70B and PaLM-540B." The paper closes the abstract with "We release all our models to the research community."

The inference-budget argument

The paper's framing argument is that the scaling-law objective of Hoffmann et al. (2022) determines how to scale dataset and model size for a given training compute budget, but "disregards the inference budget, which becomes critical when serving a language model at scale." Its statement of the alternative: "given a target level of performance, the preferred model is not the fastest to train but the fastest at inference, and although it may be cheaper to train a large model to reach a certain level of performance, a smaller one trained longer will ultimately be cheaper at inference." The authors report an empirical basis for training past the Chinchilla-optimal point: "although Hoffmann et al. (2022) recommends training a 10B model on 200B tokens, we find that the performance of a 7B model continues to improve even after 1T tokens."

The stated purpose of the release is access: "We believe that this model will help democratize the access and study of LLMs, since it can be run on a single GPU."

Training data

The paper reports a training mixture drawn only from publicly available sources compatible with open-sourcing, totalling roughly 1.4T tokens after tokenization, with each token used once except for Wikipedia and Books, over which approximately two epochs are performed.

DatasetSampling proportionEpochsDisk size
CommonCrawl67.0%1.103.3 TB
C415.0%1.06783 GB
Github4.5%0.64328 GB
Wikipedia4.5%2.4583 GB
Books4.5%2.2385 GB
ArXiv2.5%1.0692 GB
StackExchange2.0%1.0378 GB

The paper describes the preprocessing for each: five CommonCrawl dumps from 2017 to 2020 through the CCNet pipeline with line-level deduplication, fastText language identification, n-gram quality filtering and a linear classifier that discards pages not classified as Wikipedia references; GitHub filtered to Apache, BSD and MIT licensed projects; Wikipedia dumps from June–August 2022 across 20 languages in Latin or Cyrillic script; and books from Project Gutenberg together with "the Books3 section of ThePile (Gao et al., 2020)," deduplicated at book level for more than 90% content overlap. The inclusion of Books3 is the detail most often cited in the later copyright litigation over training corpora; see AI Copyright. Tokenization uses byte-pair encoding via SentencePiece, splitting all numbers into individual digits and falling back to bytes for unknown UTF-8 characters.

Architecture and compute

The models are transformers with three named modifications, each attributed to prior work: pre-normalization using RMSNorm, taken from GPT-3, normalizing the input rather than the output of each sub-layer for training stability; the SwiGLU activation function from PaLM, at dimension 2/3 · 4d rather than PaLM's 4d; and rotary positional embeddings (RoPE) from GPT-Neo, replacing absolute positional embeddings at each layer. Training uses AdamW with β₁ = 0.9, β₂ = 0.95, a cosine schedule to a final rate of 10% of maximum, weight decay 0.1, gradient clipping 1.0, and 2,000 warmup steps.

ParamsDimensionHeadsLayersLearning rateBatch sizeTokens
6.7B409632323.0e−44M1.0T
13.0B512040403.0e−44M1.0T
32.5B665652601.5e−44M1.4T
65.2B819264801.5e−44M1.4T

The authors estimate they used 2,048 A100-80GB GPUs over approximately five months to develop the models, which under their assumptions would have cost around 2,638 MWh and produced a total emission of 1,015 tCO₂eq, computed using the US national average carbon-intensity factor of 0.385 kg CO₂eq/KWh rather than any particular data center's grid.

Reported results

On zero-shot common-sense reasoning the paper reports LLaMA-65B outperforming Chinchilla-70B on all reported benchmarks but BoolQ, and PaLM-540B everywhere but BoolQ and WinoGrande, with LLaMA-13B outperforming GPT-3 on most benchmarks "despite being 10× smaller."

ModelBoolQPIQASIQAHellaSwagWinoGrandeARC-eARC-cOBQA
GPT-3 175B60.581.078.970.268.851.457.6
Gopher 280B79.381.850.679.270.1
Chinchilla 70B83.781.851.380.874.9
PaLM 540B88.082.383.481.176.653.053.4
LLaMA 7B76.579.848.976.170.172.847.657.2
LLaMA 13B78.180.150.479.273.074.852.756.4
LLaMA 33B83.182.350.482.876.080.057.858.6
LLaMA 65B85.382.852.384.277.078.956.060.2

On closed-book question answering the paper reports LLaMA-65B achieving state-of-the-art performance in zero-shot and few-shot settings on both Natural Questions and TriviaQA, and notes that LLaMA-13B is competitive on these benchmarks with GPT-3 and Chinchilla "despite being 5-10× smaller. This model runs on a single V100 GPU during inference." On TriviaQA exact match, LLaMA-65B scores 68.2 zero-shot and 73.0 at 64 shots, against Chinchilla-70B's 55.4 and 64.6 and Gopher-280B's 43.5 and 57.2.

On MMLU at five shots the paper reports LLaMA-65B at 63.4 average, "behind both Chinchilla-70B and PaLM-540B by a few percent in average, and across most domains." On mathematical reasoning it reports LLaMA-65B outperforming Minerva-62B on GSM8k "although it has not been fine-tuned on mathematical data." A single instruction-finetuning experiment following the protocol of Chung et al. (2022) produced LLaMA-I, reaching 68.9% on MMLU — better than instruction-finetuned models of moderate size, but the paper states this remains "far from the state-of-the-art, that is 77.4 for GPT code-davinci-002."

The paper also reports bias, toxicity and misinformation evaluations, including WinoGender co-reference accuracy — where the authors observe better performance on "their/them/someone" pronouns than on gendered pronouns and state this is "likely indicative of biases" — and TruthfulQA, where the 65B model reports 0.57 truthful and 0.53 truthful-and-informative.

Provenance

The arXiv abstract page and the full 27-page PDF were retrieved August 1, 2026 by direct HTTPS fetch (HTTP 200); PDF text was extracted with pdfplumber. Firecrawl was out of credits during this cycle. The capture covers pages 1–12, through the main results and the responsible-AI evaluations. The extraction does not preserve inter-word spacing in the two-column body text or table layout, and omits figures; tables above were transcribed from the extracted rows. Only v1 (submitted 27 Feb 2023) is listed on the abstract page.

Relationships

  • related: Meta AI — the publisher.
  • related: LLaMA 1 (Meta AI) — the models the paper introduces.
  • depends-on: Scaling Laws — the Hoffmann et al. result the paper's framing argument modifies by adding the inference budget.
  • related: Open-Weight Frontier Models — the release practice this paper's family is the origin case for.
  • related: AI Copyright — the Books3 and CommonCrawl composition later at issue in copyright litigation.
  • related: Llama 3 (Meta) — a later generation of the same family.