Pith. sign in

REVIEW 6 major objections 5 minor 16 references

Slimming Down LLMs Without Losing Their Minds

T0 review · 6 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read QLoRA fine-tuning on generic instruction data nearly erases a small LLM's mathematical reasoning, cuts its domain knowledge, and leaves commonsense intact.

desk verdict Small replication study with a plausible forgetting result, sunk by an unverifiable evaluation and an abstract that contradicts its own table. read the letter →

arxiv 2506.10885 v1 pith:7T2WGRTD submitted 2025-06-12 cs.CL cs.AI

Qingda (Michael) Mai This is my paper
classification cs.CLcs.AI
keywords fine-tuningQLoRALocatastrophicforgettingmathematicalreasoningHellaSwagGSM8KMMLU-CS
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show what a small language model loses when it is fine-tuned cheaply with QLoRA on a generic instruction-following dataset. Using Llama 3.2 1B and TinyLlama 1.1B, it measures commonsense reasoning (HellaSwag), mathematical reasoning (GSM8K), and computer-science knowledge (MMLU-CS) before and after fine-tuning. Its central finding is that this efficient fine-tuning nearly destroys mathematical reasoning—GSM8K flexible accuracy falls from 33.51% to 3.71%, a forgetting rate of 88.92%—and cuts MMLU-CS accuracy by 13.00 percentage points, while HellaSwag normalized accuracy changes by only +0.43%. The paper reads this as evidence that alignment between fine-tuning data and target task is decisive, and that representational competition means capabilities gained on instruction-following come at the expense of specialized reasoning and knowledge.

What carries the argument

The machinery is QLoRA's low-rank update: the base weight matrix $W$ is frozen (and 4-bit quantized), and training only touches two low-rank matrices $A \in R^{d \times r}$ and $B \in R^{r \times d}$, so the forward pass becomes $y = (W + AB)x$. This is the intervention whose effects the paper measures. The quantitative claims ride on three operationalized metrics: the Forgetting Rate $FR = (1 - A_{\mathrm{FT}}^{\mathrm{flex}}/A_{\mathrm{base}}^{\mathrm{flex}}) \times 100$, the Knowledge Loss $\Delta K = A_{\mathrm{base}} - A_{\mathrm{FT}} \pm \sqrt{CI_{\mathrm{base}}^2 + CI_{\mathrm{FT}}^2}$, and the Ability Augmentation $\Delta A = A_{\mathrm{FT}}^{\mathrm{norm}} - A_{\mathrm{base}}^{\mathrm{norm}}$. Because so few parameters are updated, the paper's argument is that any capability collapse must come from the interaction between the Alpaca fine-tuning data and the existing weights, not from wholesale rewriting of the model.

What would settle it

Re-run the paper's fine-tuning and evaluation pipeline and inspect the GSM8K outputs: if flexible extraction changes no answers on the base model (strict equals flexible accuracy), the parser is inert and the forgetting rate must be recalculated. Alternatively, check whether the fine-tuned model's 3.71% flexible accuracy comes from genuinely wrong arithmetic or from answers the extraction step fails to parse.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that QLoRA fine-tuning of Llama 3.2 1B on 2,000 Alpaca instruction samples causes catastrophic forgetting in mathematical reasoning and substantial domain-knowledge loss while leaving commonsense reasoning nearly untouched. GSM8K flexible accuracy drops from 33.51% to 3.71%, producing a forgetting rate of 88.92%, and strict accuracy drops from 33.51% to 3.11%. MMLU computer-science accuracy falls from 47% to 34%, a 13.00-point knowledge loss that the paper calls significant but partial. In contrast, HellaSwag normalized accuracy moves from 60.77% to 61.20%, a change of +0.43%. The paper also reports that TinyLlama 1.1B, fine-tuned on 50,000 Alpaca samples, scores only 27% on MMLU-CS, below the 34% of fine-tuned Llama 3.2 despite twenty-five times more training data, and interprets the overall pattern as evidence of representational competition in a fixed parameter budget.

Load-bearing premise

The headline numbers are only trustworthy if the evaluation harness and its answer-extraction and normalization rules measure the models correctly; if the GSM8K extraction rule is broken, the 88.92% forgetting rate could be an artifact.

Editorial extensions

If this is right

  • Fine-tuning a small LLM on generic instruction data without task-specific examples can erase a specialized capability almost entirely, even when only a low-rank adapter is trained.
  • Benchmarking a fine-tuned model only on the target task will miss these losses; out-of-domain evaluations on math and knowledge are needed to detect forgetting.
  • A stronger base architecture can beat a much larger fine-tuning dataset: Llama 3.2 1B's 34% MMLU-CS with 2,000 samples exceeds TinyLlama's 27% with 50,000 samples.
  • Commonsense reasoning appears more robust to this kind of instruction fine-tuning than mathematical reasoning or factual knowledge, so capability retention is domain-dependent.
  • The reported 88.92% forgetting rate gives a concrete, quantitative case of catastrophic forgetting in parameter-efficient fine-tuning, not just in full fine-tuning.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural extension the paper does not run is varying rank, quantization, and dataset mixture; mapping exactly when the GSM8K forgetting rate crosses the 50% threshold would give practitioners a rule of thumb.
  • The representational-competition explanation predicts that adding even a small block of math word problems to the Alpaca mix should cut the forgetting rate sharply; that is a cheap experiment to run.
  • Probing the adapter weights to see whether they overwrite directions in the attention projections that matter for arithmetic would test whether the observed forgetting is a parameter-interference effect.
  • Because architecture and data quantity are confounded in the TinyLlama comparison, a clean test would fix one model and vary only the fine-tuning dataset size to isolate the data effect.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

6 major / 5 minor

Summary. The paper reports an empirical study of QLoRA fine-tuning on two small language models, Llama 3.2 1B and TinyLlama 1.1B, using the Alpaca instruction-following dataset. It evaluates the base and fine-tuned models on HellaSwag, GSM8K, and MMLU-CS, and reports three headline findings: an 88.92% Forgetting Rate on GSM8K mathematical reasoning for Llama 3.2 after 2,000 Alpaca samples, a 13.00% absolute loss on MMLU-CS, and a negligible +0.43% change on HellaSwag normalized accuracy. The paper also claims that LoRA-based methods improve task-specific performance and that the results provide theoretical insight into parameter-efficient fine-tuning.

Significance. If the empirical findings were reliable, the paper would provide a useful cautionary result about generic instruction fine-tuning destroying specialized capabilities, and it would add to the literature on catastrophic forgetting in parameter-efficient adaptation. The choice of two comparable 1B models and the use of standard public benchmarks are strengths. However, the manuscript's own data contradict its abstract claim of improved task-specific performance, and the central quantitative claims rest on an evaluation protocol that is not described, whose stated metric formulas are inconsistent with the cited harness, and for which no code, prompts, or hyperparameters are provided. The theoretical contribution is limited to standard LoRA definitions and does not support the paper's stated objective. As presented, the results cannot be independently checked or interpreted, so the significance of the reported findings is currently unassessable.

major comments (6)
  1. [§2.4, Table 2] The paper's headline result, the 88.92% GSM8K Forgetting Rate, cannot be validated from the information given. The evaluation is attributed to lm-evaluation-harness, but the paper reports no prompt templates, few-shot examples, chat template, generation parameters, or answer-extraction rules. The metric formula in §2.4 defines flexible accuracy as the fraction of outputs with |p_i - t_i| < epsilon, which is a numeric-tolerance comparison and does not match lm-evaluation-harness's flexible-extract matching. Moreover, Table 2 reports identical base strict and flexible accuracies (33.51%), which suggests the flexible extractor is not actually recovering free-form answers. Because the entire forgetting claim depends on these numbers, the central result is not established.
  2. [§2.4, HellaSwag metric] The HellaSwag metric is mischaracterized. The paper defines normalized accuracy as string equality after lowercasing and removing punctuation, but lm-evaluation-harness's HellaSwag normalized accuracy compares per-token log-likelihoods normalized by sequence length. The reported '+0.43% Ability Augmentation' is therefore not the metric the authors claim to have computed. If a custom normalization was used, it must be described precisely and the associated code provided; otherwise the conclusion that commonsense reasoning is preserved is unsupported.
  3. [Abstract and §3] The abstract states that 'LoRA-based methods effectively improve task-specific performance,' but Table 2 shows no improvement on any of the three reported tasks for the fine-tuned Llama 3.2 model: GSM8K falls from 33.51% to 3.71%, MMLU-CS falls from 47% to 34%, and HellaSwag changes by only +0.43%, which is within the confidence interval. TinyLlama has no base-model baseline in the table, so no task-specific improvement can be claimed for it either. This claim is contradicted by the paper's own data and should be removed or substantially reframed.
  4. [§2.3, Table 1, Eq. (1.2)–(1.3)] The training configuration is underreported to the point that the experiments are not reproducible. The paper does not list the LoRA rank r, scaling alpha, learning rate, number of epochs or steps, batch size, target modules, quantization settings, or random seed. The mathematical formulation in §1.2.2 and §2.1.1 is only the standard LoRA definition, and no analysis links these hyperparameters to the observed results. Consequently, the stated goal of providing 'theoretical insights into parameter-efficient mechanisms' is not achieved by the current text.
  5. [§3, statistical significance] The paper reports 'p < 0.05 via paired t-test' without any details of the test: no t-statistic, degrees of freedom, number of paired items, or basis for pairing. The reported confidence intervals are consistent with binomial standard errors, but no formula or sample-size information is given. This makes it impossible to assess the statistical strength of the claimed differences, particularly for the MMLU-CS result where the 95% intervals of base and fine-tuned models do not overlap but the stated 'Knowledge Loss' includes a propagated uncertainty of ±6.93%.
  6. [Footnote 7 and availability] The paper's footnote 7 reads 'Implementation Details:' and then provides no implementation details. No code, configuration files, evaluation scripts, or model outputs are provided. For an empirical study whose central claims depend on benchmark evaluation, this omission is a load-bearing reproducibility failure. The authors should provide the exact evaluation commands, prompts, and a link to the code before the claims can be taken seriously.
minor comments (5)
  1. [Footnote 4] The aside about Stata's default output precision is irrelevant to the paper's discussion of floating-point precision and should be removed.
  2. [Throughout] There are numerous typos and grammatical errors, including 'tinnning', 'mdoel', 'downloaed', and 'lowerer'. The manuscript would benefit from careful proofreading.
  3. [References] Reference [14] is cited for the Transformer architecture but is titled 'Attention is all you need in speech separation' and does not appear to be the original Transformer paper; the standard Vaswani et al. reference should be used instead.
  4. [§2.3] The text inconsistently refers to 'Llama 1.1 1B' when it means TinyLlama 1.1B, and the model-selection justification contains redundant and somewhat contradictory claims about 'performance ceiling' and 'efficiency frontier'; these should be clarified.
  5. [Table 2 footnote 4] The table note '- indicates metric not applicable/comparable due to architectural differences' is unclear because the dash is also used for the missing TinyLlama baseline; please define exactly which entries are not applicable and why.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's claims are empirical benchmark measurements, not predictions derived from fitted inputs or self-citations.

full rationale

The paper makes no theoretical derivation that could reduce to its inputs. Its central claims—catastrophic forgetting in GSM8K (33.51% to 3.71%, FR=88.92%), preserved HellaSwag accuracy, and MMLU-CS degradation—are computed by evaluating base and fine-tuned checkpoints against fixed external benchmarks using lm-evaluation-harness. The metric formulas in Section 2.4 (Anorm, Aflex, strict/flexible accuracy, Forgetting Rate, Knowledge Loss) are measurement definitions, not fitted parameters; no quantity is fit to a subset of the benchmark results and then re-predicted. The 88.92% Forgetting Rate is arithmetic from two independently measured accuracies. All citations are to external methodology, datasets, and frameworks (LoRA, QLoRA, Alpaca, lm-evaluation-harness); none encodes the paper's conclusion, and the paper contains no self-citations or imported uniqueness theorems. The absent 'Implementation Details' and the potential prompt/extraction mismatch flagged by the skeptic are reproducibility and validity concerns, not circularity, because the evaluation protocol is not constructed from the conclusion it is used to support. No circular step can be exhibited from the paper's own equations or citation chain.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The empirical claims depend on unreported hyperparameters and the correct functioning of evaluation tools; no new theoretical quantities or entities are introduced.

free parameters (5)
  • LoRA rank r = not reported
    The paper states 'typically r=8' but never reports the value used; the number of trainable parameters and the adaptation capacity depend directly on r.
  • LoRA scaling alpha = not reported
    Scaling factor for LoRA updates; affects the magnitude of weight modifications and thus the degree of forgetting.
  • Learning rate = not reported
    Affects convergence and stability; not disclosed.
  • Number of training epochs (or steps) = not reported
    The extent of fine-tuning determines how much the model shifts from its pretrained state.
  • GSM8K tolerance threshold epsilon = 0.01
    Used in the flexible accuracy formula; the choice influences the flexibility measure but is standard.
assumptions (3)
  • domain assumption lm-evaluation-harness computes benchmark accuracies correctly, and the paper's metric formulas (flexible accuracy, normalization, knowledge loss) are implemented as intended.
    Section 2.4 invokes the harness for all scores; if the implementation has bugs (e.g., producing identical strict and flexible GSM8K accuracy), the central numbers are unreliable.
  • domain assumption The Alpaca dataset is a representative instruction-tuning corpus for testing parameter-efficient fine-tuning.
    Section 2.3 selects Alpaca as the fine-tuning data; the observed forgetting pattern is a property of this choice, not necessarily of QLoRA in general.
  • domain assumption The QLoRA implementation in Unsloth preserves the pretrained weights and correctly merges adapters after training.
    Section 2.2.3 describes the merge process; if the merge or quantization is flawed, the post-fine-tuning evaluations do not reflect the intended model state.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Slimming Down LLMs Without Losing Their Minds." pith.science (2026). https://pith.science/paper/7T2WGRTD

@misc{pith2026250610885,
  author       = {Pith},
  title        = {Pith review of: Slimming Down LLMs Without Losing Their Minds},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7T2WGRTD}},
  note         = {Machine review of arXiv:2506.10885}
}
read the original abstract

This paper investigates and validates the impact of fine-tuning on large language model performance, focusing on parameter-efficient methods (LoRA and QLoRA). We evaluate model capabilities across three key domains: (1) commonsense reasoning (HellaSwag), (2) mathematical reasoning (GSM8K), and (3) multi-domain knowledge (MMLU-CS). Our findings demonstrate that: (1) LoRA-based methods effectively improve task-specific performance while maintaining computational efficiency, and (2) performance strongly depends on alignment between fine-tuning dataset and benchmark tasks. The study provides both theoretical insights into parameter-efficient mechanisms and practical guidance for developers implementing efficient LLM adaptation with limited resources.

Figures

Figures reproduced from arXiv: 2506.10885 by the authors.

Figure 1
Figure 1. Transformer architecture with parameter-efficient tuning methods. Dashed blocks represent added modules: (1) LoRA updates weight matrices through low-rank decomposition, (2) Adapters insert projection layers, (3) Prefix Tuning prepends learnable vectors to attention keys/values. Picture Source from [8] 2.2.1 Recap for Transformer Base Architecture • Multi-Head Attention [6, 14]: Computes contextual relationships via… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

16 extracted references · 1 canonical work pages

  1. [1]

    Aghajanyan, L

    A. Aghajanyan, L. Zettlemoyer, and S. Gupta, Intrinsic dimensionality explains the effectiveness of language model fine-tuning, 2020, http://dx.doi.org/10.48550/arxiv.2012.13255

    arXiv doi:10.48550/arxiv.2012.13255 2020

  2. [2]

    Cobbe, V

    K. Cobbe, V . Kosaraju, M. Bavarian, M. Chen, H. Jun, L. Kaiser, M. Plappert, J. Tworek, J. Hilton, R. Nakano, C. Hesse, and J. Schulman, Training verifiers to solve math word problems, 2021, http://dx.doi.org/10.48550/arxiv.2110.14168

    arXiv doi:10.48550/arxiv.2110.14168 2021

  3. [3]

    Dettmers, M

    T. Dettmers, M. Lewis, S. Shleifer, and L. Zettlemoyer, 8-bit optimizers via block-wise quantization, 2021, http://dx.doi.org/10.48550/arxiv.2110.02861

    arXiv doi:10.48550/arxiv.2110.02861 2021

  4. [4]

    Dettmers, A

    T. Dettmers, A. Pagnoni, A. Holtzman, and L. Zettlemoyer, QLoRA: Efficient Finetuning of Quantized LLMs, 2023, http://dx.doi.org/10.48550/arxiv.2305.14314

    arXiv doi:10.48550/arxiv.2305.14314 2023

  5. [5]

    Dettmers and L

    T. Dettmers and L. Zettlemoyer, The case for 4-bit precision: k-bit Inference Scaling Laws, 2022, http://dx.doi.org/10.48550/arxiv.2212.09720

    arXiv doi:10.48550/arxiv.2212.09720 2022

  6. [6]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova, Bert: Pre-training of deep bidirectional transformers for language understanding, arXiv preprint arXiv:1810.04805 , 2018, http://dx.doi.org/10.48550/arXiv.1810.04805. 10

    arXiv doi:10.48550/arxiv.1810.04805 2018

  7. [7]

    Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac’h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang Sutawika, Eric Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou, A framework...

    arXiv 2024

  8. [8]

    J. He, C. Zhou, X. Ma, T. Berg-Kirkpatrick, and G. Neubig, Towards a unified view of parameter-efficient transfer learning, 2021, http://dx.doi.org/10.48550/arxiv.2110.04366

    arXiv doi:10.48550/arxiv.2110.04366 2021

Show all 16 references
  1. [9]

    Hendrycks, C

    D. Hendrycks, C. Burns, S. Basart, A. Zou, M. Mazeika, D. Song, and J. Steinhardt, Measuring massive multitask language understanding, 2020, http://dx.doi.org/10.48550/arxiv.2009.03300

    2020 doi arXiv

  2. [10]

    Houlsby, A

    N. Houlsby, A. Giurgiu, S. Jastrzebski, B. Morrone, Q. de Laroussilhe, A. Gesmundo, M. Attariyan, and S. Gelly, Parameter-efficient transfer learning for NLP, 2019, http://dx.doi.org/10.48550/arxiv.1902.00751

    2019 doi arXiv

  3. [11]

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, LoRA: Low-Rank Adaptation of Large Language Models, 2021, http://dx.doi.org/10.48550/arxiv.2106.09685

    2021 doi arXiv

  4. [12]

    X. L. Li and P. Liang, Prefix-tuning: Optimizing continuous prompts for generation, 2021, http://dx.doi.org/10.48550/arxiv.2101.00190

    2021 doi arXiv

  5. [13]

    S. J. Pan and Q. Yang, A survey on transfer learning, IEEE Transactions on Knowledge and Data Engineering, 22(10):1345–1359, 2010, http://dx.doi.org/10.1109/TKDE.2009.191

    2010 doi

  6. [14]

    Subakan, M

    C. Subakan, M. Ravanelli, S. Cornell, M. Bronzi, and J. Zhong, Attention is all you need in speech separation, pp. 21–25, 2021, http://dx.doi.org/10.1109/ICASSP39728.2021.9413901

    2021

  7. [15]

    N. Wang, J. Choi, D. Brand, C.-Y . Chen, and K. Gopalakrishnan, Training deep neural networks with 8-bit floating point numbers, 2018, http://dx.doi.org/10.48550/arxiv.1812.08011

    2018 doi arXiv

  8. [16]

    Zellers, A

    R. Zellers, A. Holtzman, Y . Bisk, A. Farhadi, and Y . Choi, HellaSwag: Can a Machine Really Finish Your Sentence?, 2019, http://dx.doi.org/10.48550/arxiv.1905.07830. 11

    2019 doi arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.