yoxia/FocusOYL-Prism-1B

FocusOYL Prism-1B

Research repository: FocusOYL MathMorph · Technical report · Experiment history · Evaluation notes

Training-free spectral editing of MiniCPM5-1B.

FocusOYL Prism-1B is an experimental, 1B-class language model derived from OpenBMB's MiniCPM5-1B. It explores whether targeted mathematical edits to existing weights can improve task performance without additional training or changes to the inference architecture.

This release applies a partial spectral projection to a single feed-forward weight matrix and distributes the resulting checkpoint as F16 GGUF. It is a weight-edited derivative, not a foundation model pretrained from scratch.

Release status: Experimental; full local GSM8K + IFEval comparison completed. Under matched Think-on settings, Prism reaches 72.71% GSM8K flexible-extraction accuracy versus 70.43% for the original, while IFEval prompt-level strict accuracy declines to 68.58% from 72.64%. These are self-reported public-benchmark results, not official leaderboard scores or an across-the-board upgrade.

Mathematical specification of the released checkpoint

This checkpoint uses ordinary Euclidean partial spectral projection. It does not use the high-dimensional feature metric or readout-protected variants explored elsewhere in the research.

For the original tensor $WW$ named blk.23.ffn_down.weight, let $W=U\mathrm{\Sigma }{V}^{\mathrm{\top }}W=U\backslash Sigma\; V^\backslash top$ and $P_{512}=U_{512}{U}_{512}^{\mathrm{\top }}P\_\{512\}=U\_\{512\}U\_\{512\}^\backslash top$. The residual is $D=(I-P_{512})WD=(I-P\_\{512\})W$.

$$\boxed{{\displaystyle W^{\mathrm{\prime }}=W-\alpha D=P_{512}W+(1-\alpha )(I-P_{512})W}}\backslash boxed\{W\text{'}=W-\backslash alpha\; D=P\_\{512\}W+(1-\backslash alpha)(I-P\_\{512\})W\}$$

$$\alpha =\min (1,\frac{0.98\times 0.12\parallel W{\parallel }_F}{\parallel D{\parallel }_F})\mathrm{.}\backslash alpha=\backslash min\backslash left(1,\backslash frac\{0.98\backslash times0.12\backslash thinspace\; \backslash lVert\; W\backslash rVert\_F\}\{\backslash lVert\; D\backslash rVert\_F\}\backslash right).$$

The zero-residual case is unchanged. In exact arithmetic,

$$\frac{\parallel W^{\mathrm{\prime }}-W{\parallel }_F}{\parallel W{\parallel }_F}\le 0.1176,\parallel W^{\mathrm{\prime }}{\parallel }_F^2=\parallel P_{512}W{\parallel }_F^2+(1-\alpha {)}^2\parallel D{\parallel }_F^2\mathrm{.}\backslash frac\{\backslash lVert\; W\text{'}-W\backslash rVert\_F\}\{\backslash lVert\; W\backslash rVert\_F\}\backslash leq0.1176,\; \backslash qquad\; \backslash lVert\; W\text{'}\backslash rVert\_F^2=\backslash lVert\; P\_\{512\}W\backslash rVert\_F^2+(1-\backslash alpha)^2\backslash lVert\; D\backslash rVert\_F^2.$$

Only this tensor is edited. The projector dimension is not the rank of the entire model. F16 rounding is checked after writing; the bound is a weight-space statement, not a guarantee of more accurate answers or preserved stopping behavior.

The technical report derives the earlier GPEP/NPSR operators, row-tangent updates, cubic Hermite interaction geometry, downstream metrics, and readout-contrast protection. Those are separate experiments, with their negative results retained in the experiment log. The completed full-split GSM8K and IFEval comparison is documented in benchmark results and evaluation notes. The earlier internal study is preserved separately, not overwritten or relabeled as a leaderboard result.

Model details

Property	Value
Model	FocusOYL Prism-1B
Release	F16-Experimental-01
Publisher	yoxia
Base model	OpenBMB / MiniCPM5-1B
File	FocusOYL-Prism-1B-F16-Experimental-01.gguf
Format and size	F16 GGUF; 2,166,551,936 bytes (approximately 2.17 GB)
Editing recipe	MathMorph v0.4 / euclid_control
Additional training	None: no gradient updates, LoRA, distillation, or teacher model
Inference architecture	Unchanged; no added layers, parameter slots, RAG, or external agent
License	Apache-2.0

Prism is the Euclidean projection variant. It is not the earlier NPSR sharpen checkpoint or one of the cubic-feature geometry candidates. Results from those variants must not be attributed to this release.

Method

Starting from the unmodified MiniCPM5-1B F16 GGUF, the conversion edits only blk.23.ffn_down.weight (zero-based block index).

The method identifies a leading left singular subspace with a projection rank of 512, partially shrinks the complementary component, and writes the edited matrix back in its original F16 representation. The per-tensor relative-change budget is 0.12. Rank 512 describes the projection subspace, not the rank or parameter count of the entire model.

Local byte-level audits confirmed that all other tensors, the tokenizer, the embedded chat template, the architecture, and the file size were preserved. See MODIFICATIONS.md for the change notice.

Evaluation questions and answer labels did not enter the weight computation. Development-set scores were used in the broader research selection process. Training-free therefore means no additional training of this checkpoint; it does not mean that the base model was never trained or that the research workflow was entirely data-independent.

Running the model

Load the GGUF file in a compatible llama.cpp build or another GGUF runtime whose compatibility you have verified. Use the original chat template embedded in the file. This repository is not a Transformers safetensors checkpoint directory.

The base model supports Think and No Think modes. Configure the mode according to the upstream model card and your runtime's documentation, and verify that it actually takes effect. Upstream capability descriptions are not evidence that this edited checkpoint has passed the same evaluations.

For a fair comparison, use the original F16 baseline and this exact Prism release with matching runtime, chat template, reasoning mode, sampling settings, generation budget, and hardware conditions.

Evaluation

Completed public-benchmark comparison

Completed local, full-split evaluation using public benchmark task definitions. Both models use Think mode, zero-shot prompts, greedy decoding, and a common 4,096-token total generation cap. This is a self-reported experiment, not an independently verified leaderboard submission.

Benchmark / metric	MiniCPM5-1B F16	FocusOYL Prism-1B F16	Change (pp)
GSM8K: flexible numeric extraction	70.43% (929/1,319)	72.71% (959/1,319)	+2.27
GSM8K: strict-format extraction	0.23% (3/1,319)	0.53% (7/1,319)	+0.30
IFEval: prompt-level strict	72.64% (393/541)	68.58% (371/541)	-4.07
IFEval: instruction-level strict	74.70% (623/834)	71.34% (595/834)	-3.36
IFEval: prompt-level loose	74.12% (401/541)	71.35% (386/541)	-2.77
IFEval: instruction-level loose	75.78% (632/834)	73.50% (613/834)	-2.28

Prism answers 30 more GSM8K questions correctly under flexible numeric extraction, but passes all strict IFEval instructions on 22 fewer prompts. All four IFEval metrics decline. This is a task trade-off, not an across-the-board capability upgrade.

Each model was evaluated on 1,319 GSM8K questions and 541 IFEval prompts. IFEval contains 834 individual instruction checks; those are not 834 separate prompts. Changes are calculated from unrounded scores.

GSM8K strict-format extraction is a formatting-sensitive diagnostic; both extraction filters are retained. Full benchmark report, token usage, extraction notes, and audit.

Internal controlled study

The following results come from a fixed held-out study conducted on September 5, 2026. Four duplicate prompts were identified and excluded from certification before held-out evaluation began, leaving 156 non-overlapping test items. All 160 original records were retained. The table reports only Prism and its original F16 baseline.

Test subset	Items	Original F16: correct	Prism-1B: correct
ARC-Challenge subset	64	38	48
ARC-Easy subset	64	49	50
Synthetic arithmetic	16	15	16
Synthetic Python code interpretation	12	8	7
Total	156	110 (70.51%)	121 (77.56%)

Prism delivered 11 more correct answers, an observed increase of 7.05 percentage points. Total generated tokens decreased from 107,200 to 97,038, approximately 9.48%. Seventeen baseline failures became successes, while six baseline successes became failures.

The two-sided paired test gave an unadjusted p = 0.03469 and p = 0.06938 after correction for two comparisons. The code-interpretation subset lost one correct answer. These results did not satisfy the predefined overall acceptance criteria and do not establish a reliable, across-the-board improvement over the original model.

This is a custom subset evaluation, not a complete official ARC benchmark or an independently verified leaderboard submission. The evaluation summary records the study's aggregate results; it does not contain the complete raw responses or constitute third-party verification.

Internal study settings

Setting	Value
Reasoning mode	Think enabled for both models
Sampling	temperature=0, top_p=1, top_k=0, min_p=0, repeat_penalty=1
Generation budget	2,048 tokens per item, including reasoning and the final answer
Context window	4,096 tokens per request; context shifting disabled
Concurrency	Four requests
Observation guard	120 seconds per item
Runtime	llama.cpp b10672 / 511f9c137
Success criterion	A correct, parseable final answer after natural completion

Truncation, parsing failures, and loop-guard interruptions did not count as successes. Reasoning text was not searched for a correct answer to substitute for the final response. The loop guard belonged to the evaluator, not the GGUF model.

Public science questions came from AllenAI ai2_arc, pinned to revision 210d026faf9955653af8916fad021475a3f00453. Arithmetic and code items were generated for the study. Overlap with knowledge or patterns in pretraining cannot be ruled out. This card does not reproduce the dataset's questions.

Reproducibility and external evaluation

External evaluations should record model-file checksums, dataset revisions, scorer versions, chat templates, reasoning mode, generation limits, sampling settings, and hardware. Report each benchmark separately, alongside actual token usage, truncations, and failures. A model upload or a self-reported score is not an independent certification or leaderboard acceptance.

File integrity

Expected SHA-256 of the Prism GGUF:

24f86e98d327afb5d17418486708b0dc653749d6eed67150c4f58f8c6591a41b

SHA-256 of the original F16 file used as the local baseline:

68c40b08b1242754a107b9510af89aa75b10c75843ca7844643c70956b7f1e3d

These checksums identify the files in the local experiment. The uploaded Hub weights were not independently re-downloaded and hashed when this documentation was prepared. Verify downloaded bytes against the expected checksum before external evaluation. Renaming a file does not change its content hash.

Intended use and limitations

This checkpoint is intended for research on training-free weight editing, small-model behavior, and inference efficiency. It is not an accepted production replacement for the baseline.

Outputs may be incorrect, biased, repetitive, or incomplete. The evidence covers one weight-edited checkpoint, two full public benchmark tasks under a fixed decoding configuration, and earlier internal subsets. It does not establish long-context reliability, tool-use performance, multimodal capability, or safety. In particular, 12 code-interpretation items do not characterize general coding ability. Do not use this experimental model as an unverified basis for high-stakes decisions.

License and attribution

MiniCPM5-1B was released by OpenBMB under Apache-2.0, as stated in its upstream model card. This derivative retains attribution to that source and uses the same license. LICENSE contains the license text; MODIFICATIONS.md documents the changes and does not replace any applicable upstream notices.

Naming the base model establishes provenance, not endorsement of this release or its results by OpenBMB, Hugging Face, or an evaluation organization.

References: MiniCPM5-1B · llama.cpp · ARC dataset

Converter entry points

The implementation is available in FocusOYL-MathMorph. Use convert_prism.py to reproduce the fixed released recipe from its exact baseline. Use convert_euclidean.py for explicit Euclidean projection experiments with configurable layers, rank, and per-tensor update budget. The experimental entry shares the unchanged v0.4 core and includes a --dry-run metadata preflight.

See the Euclidean converter guide for installation, supported storage types, safeguards, and examples. The converters do not themselves run benchmarks or certify capability improvements.