[Alajmah]

Summary

This stacked PR adds the compressed 2017–2022 foundation layer needed to explain ModelMap's high-resolution 2023–2026 history.

Historical spine

Adds canonical records and dated primary-source events for GPT-1, BERT, GPT-2, T5, GPT-3, Switch Transformer, InstructGPT, Chinchilla, and PaLM. It also records the foundational papers for neural scaling laws, RMSNorm, SwiGLU, FLAN-style instruction tuning, chain-of-thought prompting, and FlashAttention.

Technique layer

Adds canonical technique entities for generative pretraining, masked language modeling, text-to-text transfer, neural scaling laws, in-context learning, RMSNorm, SwiGLU, instruction tuning, chain-of-thought prompting, RLHF, compute-optimal training, and FlashAttention. Existing Transformer and sparse-MoE technique nodes are reused rather than duplicated.

Evidence and lineage

Every new model release has evidence-backed family and publisher edges. Model-to-technique edges are limited to claims directly supported by primary sources. InstructGPT is explicitly modeled as fine-tuned from GPT-3. GPT-2 has an explicit successor edge from the 2018 GPT model because OpenAI's 2019 announcement calls GPT-2 a successor/direct scale-up; the GPT-1 label is marked as retrospective rather than presented as 2018 branding.

Scale progression

Adds evidence-backed value claims for GPT-2 (1.5B parameters), GPT-3 (175B), Switch Transformer experiments up to one trillion parameters, Chinchilla (70B), and PaLM (540B). These remain atomic claims rather than graph edges.

Institutional history

Adds historical org:google and org:deepmind entities so pre-2023 work is not retroactively attributed to the later Google DeepMind organization.

Chronology

Extends CI year-view generation to 2017–2022, producing a continuous generated timeline from the existing Transformer event through the high-resolution 2023–2026 corpus.

Scope

This is intentionally not an exhaustive pre-2023 NLP catalog. It includes only models and ideas that materially explain the architecture, scaling, post-training, reasoning, and inference developments emphasized in 2023–2026.

Validation

Current head 0e9639d is mergeable and green in validate-corpus run #40. Primary-source dates and the high-risk historical naming/lineage decisions were manually audited. Codex code-review quota is currently exhausted, so no fresh Codex review is claimed for this head.

Stack

Base: atlas-2026-v0.1 (PR #8)
Head: atlas-foundations-2017-2022-v0.1

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.