Methods › General › Fine-Tuning › Discriminative Fine-Tuning

Discriminative Fine-Tuning

1,990 papers tagged archive 2025-07-28

Introduced by Jeremy Howard et al. in Universal Language Model Fine-tuning for Text Classification

archive 2025-07-28 Description, source and code snippet are the archive's method entry.

Discriminative Fine-Tuning is a fine-tuning strategy that is used for ULMFiT type models. Instead of using the same learning rate for all layers of the model, discriminative fine-tuning allows us to tune each layer with different learning rates. For context, the regular stochastic gradient descent (SGD) update of a model’s parameters θ at time step t looks like the following (Ruder, 2016):

θₜ = θₜ₋₁ − η·∇_θJ(θ)

where η is the learning rate and ∇_θJ(θ) is the gradient with regard to the model’s objective function. For discriminative fine-tuning, we split the parameters θ into {θ₁, …, θ_L} where θₗ contains the parameters of the model at the l-th layer and L is the number of layers of the model. Similarly, we obtain {η₁, …, η_L} where θₗ where ηₗ is the learning rate of the l-th layer. The SGD update with discriminative finetuning is then:

θₜˡ = θₜ₋₁ˡ - ηˡ·∇_(θˡ)J(θ)

The authors find that empirically it worked well to first choose the learning rate ηᴸ of the last layer by fine-tuning only the last layer and using ηˡ⁻¹=ηˡ/2.6 as the learning rate for lower layers.

PaperSourceSee Code · fastai/fastai

In Syntology for a different name, 'Discriminative Finetuning', about another paper, arXiv:2304.01662: repository known, no verified implementation (declared_code; repositories: 1 listed of 1 in the graph, strongest link: official). The lookup matched a different name in Syntology's graph, 'Discriminative Finetuning' (resolution: token_match); this answer is about that name. About another paper: this answer is about arXiv:2304.01662, while the archive credits arXiv:1801.06146.

Code Syntology holds under other names, as the lookup lists them (code_nearby); not a complete list, and none of them is this method.

Origin paper of this answer, arXiv:2304.01662 (its page on this site · its page on Syntology): Syntology's run record for its harvested code, 2 ran (of which 0 constructed an object rather than computing a result; 2 with no instrument failure: 0 honoured, 0 violated, 2 with no contract checked; 0 where Syntology's instrument failed) · 0 unverified (of 2 harvested samples). The instrument figure counts failures of Syntology's instrument, not of the code. Ran means executed on a synthesized input, not that the code is correct.

Syntology's have check for this name (free, it answers coverage only), read 2026-10-05; ask it again live. A level is the served verification level of that one answer; nothing here is summed across methods.

What each level means, quoted from Syntology's server card: V0 “imports”; V1 “executes against a synthesized fixture and returns a value of the recorded type”; V2 “two implementations GENERATED from the paper's signature and spec agree numerically (1e-6) on fixture inputs” “neither of the two ever read the authors' own code, so V2 measures how unambiguously the paper specifies the method, not fidelity to the authors' implementation (measured 2026-09-11, commit 75eaa77: cross_checked_by is a second LLM on 2,628 of 2,831 served samples)”; V3 “paper-derived property tests pass on both of those same two generations.” The card adds: “No rung on this ladder compares the served code against the authors' implementation, so treat every level as evidence about runnability and internal agreement, not about matching the paper's results.” A level quoted in two parts is one definition the card writes as two clauses joined by a double hyphen, each clause quoted on its own (how these levels are quoted).

Papers archive 2025-07-28

archive papers tagged: 1,990 · with a code link: 794 · where Syntology ran a sample: 271 (216 with a run with no instrument failure, 55 where every run was a failure of Syntology's instrument) Syntology

Show: all tagged papersonly where code ran (271 of 1,990 tagged: 216 with a run with no instrument failure, 55 where every run was a failure of Syntology's instrument)

Page 1 of 20: papers 1 to 100 of 1,990, newest first by the archive's date (ties by slug), in archive order.

Repository counts are the archive's code-links table. A Syntology line states what Syntology ran from that paper's harvested code, as “N ran (of which C constructed an object rather than computing a result; K with no instrument failure: H honoured, V violated, P with no contract checked; I where Syntology's instrument failed) · U unverified”; the instrument figure counts failures of Syntology's instrument, not of the code. It is per sample and not a correctness claim. When the archive marks a repository official for the paper, the line starts with that repository's state (the archive's flag, not a verdict on who wrote the code; “community repositories only” when every sample that ran came from a community repository, “official: no sample here; runs from other or unrecorded repositories” when some came from a repository the paper names or has in its text, or from none recorded); hover it for the repositories the samples that ran came from.

Tasks archive 2025-07-28

20 shown of 725 tasks the archive attaches to papers tagged with this method, by distinct papers. A task without a page in the catalog is plain text.

TaskPapers
Language Modelling431
Language Modeling323
Text Generation192
Large Language Model123
Question Answering106
Sentence99
Decoder81
Retrieval78
Transfer Learning65
Text Classification59
Prompt Engineering57
text-classification54
Decision Making53
In-Context Learning53
Translation53
Sentiment Analysis52
GPU49
Natural Language Understanding40
Few-Shot Learning39
Machine Translation39

Usage over time archive 2025-07-28

Papers per year tagged with Discriminative Fine-Tuning: 2018 to 2025, peak 701 701 0 2018: 2 papers 2018 2019: 68 papers 2019 2020: 141 papers 2020 2021: 173 papers 2021 2022: 164 papers 2022 2023: 491 papers 2023 2024: 701 papers 2024 2025: 250 papers 2025
Papers per year the archive tags with this method, by the paper's archive date (1,990 dated). Bars are counts, not a trend claim.

Components: the archive holds no method-to-method composition, so PwC's Components table cannot be rebuilt; the Papers list carries no Results column for the same reason (the archive does not join its leaderboard rows to method tags).

Categories archive 2025-07-28

Fine-Tuning

Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections