NotCherry/LFM3-1.7B-ThinkExit-GGUF

LFM3-1.7B ThinkExit · experimental GGUF

A small, training-free weight edit that changes when LFM3 leaves its thinking phase. The 1.7B backbone and original input embeddings are preserved. No gradient training, LoRA training, or additional knowledge training was performed.

This is a research artifact, not a proven general upgrade. The selected edit reached the requested point scores on fresh math and instruction-following subsets, but its gains over the previous edited model are statistically inconclusive and it used 29.2% more generated tokens in that comparison.

At a glance

Property	Value
Upstream architecture	LiquidAI/LFM3-1.7B
Exact source checkpoint	AtomicChat LFM3-1.7B Uncensored AD-Q4_K_M
Format	Standalone GGUF, Q4_K_M mixed quantization; copied head remains Q6_K
Edit	Explicit output head; scale only its </think> row by 1.50
Preserved	Every original tensor payload, including input embeddings and backbone
Storage tradeoff	About 215 MB extra; 262,144,000 additional stored values, mostly duplicates
Training	None
Evaluation status	Local paired evaluation; not an independent leaderboard certification

What changed

The original GGUF ties its input embedding to the output projection. This version copies that tensor into an explicit output.weight, leaves the input embedding unchanged, and scales eight Q6_K block scales in the output row for </think> (token ID 124902). Only 10 bytes in that copied output row differ from its source; all other output rows are unchanged. It adds storage, not an additional projection operation. The resulting serialized model stores approximately 1.99B values, including the duplicated head, rather than the original approximately 1.73B.

The hypothesis is completion calibration: encourage a timely transition from reasoning to a final response. It does not demonstrate new learned knowledge. The experiment was inspired by Weight-Space Programming; the article's Qwen-specific intervention was not directly applied to LFM3.

The source already contains AtomicChat's safety ablation and quantization. This release inherits those changes and does not claim to restore safety refusals. It is not an official Liquid AI or AtomicChat release.

Paired math and instruction evaluation

The selected artifact was frozen before generating any of these confirmation responses. One fixed-seed answer per question; no majority voting, test-feedback retries, or choosing the best response.

Fresh held-out subset	Previous tied 2.00 edit	This output-only 1.50 edit
GSM8K · 192 questions	175/192 · 91.15%	177/192 · 92.19%
MATH-500 · 128 questions	115/128 · 89.84%	117/128 · 91.41%
IFEval · 192 prompts, strict prompt metric	176/192 · 91.67%	178/192 · 92.71%

The comparator here is an earlier experimental edit, not the original downloaded GGUF and not the official BF16 model. Both use the same llama.cpp commit and CUDA runtime, thinking enabled, 16,384 generated tokens, temperature 0.1, top-k 50, top-p 1.0, repetition penalty 1.1, and matching per-question seeds. Thinking tokens count toward the output limit.

Metric	Observation
Net extra correct answers	2 per benchmark
Paired 95% accuracy-change interval, GSM8K	−2.6 to +4.7 percentage points
Paired 95% accuracy-change interval, MATH-500	−4.7 to +7.8 percentage points
Paired 95% accuracy-change interval, IFEval	−3.1 to +5.2 percentage points
Total generated tokens	1,105,811 → 1,428,701 · 29.2% increase

All accuracy intervals include zero. These are subset point estimates, not full benchmark results or proof of improvement. Intervals cover variation across questions, not generation seeds or GPU batches. Earlier CPU/2,048-token results also changed questions, runtime and budget, so the large historical score jump cannot be attributed to this weight edit alone.

Math scoring uses math-verify on the final response. IFEval uses the official Google checker. A narrow numeric-format parser fix was applied equally to all models before confirmation; it corrected one training-development answer key with a thousands separator. No answer label was changed.

Development used GSM8K/MATH training examples and synthetic instruction prompts. No official IFEval evaluation prompt was used for round-two selection. Stronger 3x and 4x output edits caused closing-token loops and were rejected; those failures are part of the experiment, not successful candidates.

Cross-domain check

An additional evaluation-only comparison is running on all 164 HumanEval tasks, 256 ARC-Challenge test questions, and 256 BoolQ validation passages. It compares this exact binary with the original downloaded AtomicChat GGUF under identical inference settings. Neither weights nor settings will be revised from these results. Results will be added when both models finish.

HumanEval code runs in separate rootless, nonroot containers without network or host filesystem mounts. All 164 reference solutions passed the test setup. This local isolation is not the same as an independent third-party evaluation. Public benchmark contamination during upstream pretraining is unknown.

Run locally

Use a current llama.cpp build with LFM3 and optional explicit-output-head support. The evaluated commit was fac3a894fc04901c3e5abaff457e024c6b52e04f. See pinned downloads, hashes, source build, and API instructions.

llama-server \
  -hf NotCherry/LFM3-1.7B-ThinkExit-GGUF \
  -ngl 99 -c 32768 -np 1 \
  --jinja --reasoning-format none

For matching the reported inference preset, send chat-completions requests with:

{
  "messages": [{"role": "user", "content": "Your question here"}],
  "temperature": 0.1,
  "top_k": 50,
  "top_p": 1.0,
  "repeat_penalty": 1.1,
  "max_tokens": 16384,
  "seed": 20261105,
  "chat_template_kwargs": {"enable_thinking": true}
}

The exact benchmark prompts, subset provenance, seeds, scores and selection history are described in the included evaluation records. The example seed above is illustrative; evaluations use the recorded per-question seeds.

Provenance and files

• Model: LFM3-1.7B-output-exit-1.50-Q4_K_M.gguf
• Model SHA-256: 2ff59a7013c6608fa3dc643e053964f57c9585cdd8532330d3b0583ae2628acc
• Source SHA-256: b22df65673b54ed326f18f74826d76f3b65dadc237ffd6306cd52be04544827d
• Source Hub revision: 3ad712542eb9de349dbd28f2739b69c925cdec7d
• Exact edit manifest
• Inference preset
• Paired math/instruction results
• Frozen selection and development trials
• Final response and scoring audit
• Cross-domain protocol
• Cross-domain dataset revisions

Release metadata normalizes local filesystem paths. The GGUF binary is unchanged from the evaluated artifact; its SHA-256 is the authoritative identifier.

Limits and license

This is a narrow termination edit. It can regress individual answers, formatting, multiturn behavior, or other capabilities. It is not validated as a software engineering agent, and simple function tests do not establish SWE-bench ability. The published model has no independent Arena rank or private-test certification.

The original LFM Open License v1.0 applies, including its commercial-use conditions. See NOTICE for origin and modification notices. Original model authors and AtomicChat retain their respective rights. This release does not replace their license with a different license.