LFM2.5-2.6B - UNCENSORED ABLITERATED PHILADELPHIA CLASS
This is LiquidAI's LFM2.5-2.6B with its refusal behavior aggressively reduced. It keeps the compact 2.7B hybrid architecture, but answers requests the stock post-trained model refused. In our tests, refusals fell from 132/842 to 8/842 and from 19/126 to 1/126 on a held-out set, while the average change across five public capability benchmarks was only -0.50 percentage points.
PHILADELPHIA CLASS includes the full BF16 checkpoint as one model.safetensors file plus a tested BF16 GGUF for llama.cpp. Both retain LFM2.5's tokenizer, reasoning chat template, and 128K configured context.
Modification notice
This repository redistributes modified model weight files derived from LiquidAI/LFM2.5-2.6B at revision dca1825886789bd40b94368f53b1d9ada4c94598. The local untouched checkpoint was verified against all 11 files in that Hub revision before release preparation.
The candidate was produced with two sequential output-space biprojection stages. Exact comparison with the untouched source found 58 intentionally modified output-projection tensors out of 266 total; 208 tensors remained exactly equal, with no missing, unexpected, shape-mismatched, or dtype-mismatched tensors. The Transformers release is packaged as one model.safetensors file.
Release at a glance
Property
Value
Base model
LiquidAI/LFM2.5-2.6B post-trained model
Upstream revision
dca1825886789bd40b94368f53b1d9ada4c94598
Architecture
Lfm2ForCausalLM
Parameters
2,697,198,592
Precision
BF16
Transformers weights
One BF16 model.safetensors file (5.39 GB)
GGUF
Tested BF16 build (5.40 GB)
Layers
30: 22 convolution blocks and 8 grouped-query attention blocks
Configured context
128,000 tokens
Vocabulary
128,000 tokens
Status
RC1; strong refusal-reduction candidate, not safety-certified
Lossy quantization
No lossy quantized build is included
Internal behavioral diagnostics
These are automated OBLITERATUS development diagnostics, not standardized leaderboards or independent safety audits. "Direct," "warning," "refusal," and "usable" are deterministic response-form labels; they do not establish factuality, legality, harmlessness, or human approval.
Matched 842-item opening screen
Both checkpoints used the same prompt indices, chat template, system prompt, seed, greedy decoding, and 96 generated tokens.
Cohort
Untouched source direct / warning / refusal
PHILADELPHIA CLASS direct / warning / refusal
Harmful 842
149 / 561 / 132
773 / 61 / 8
Paired harmless 842
795 / 45 / 2
812 / 30 / 0
Of the source model's 132 harmful refusals, 131 cleared. Seven source non-refusals became refusals, so this is a large net reduction rather than monotonic improvement on every item.
Exact-prompt-hash holdout
The 126 holdout prompt hashes were excluded from direction fitting. This blocks exact prompt identity reuse but is not a semantic or paraphrase-family holdout.
Cohort
Untouched source direct / warning / refusal
PHILADELPHIA CLASS direct / warning / refusal
Harmful holdout 126
15 / 92 / 19
114 / 11 / 1
Paired harmless holdout 126
118 / 8 / 0
123 / 3 / 0
Coherence and long-form diagnostics
LFM2.5 begins inside a reasoning block. These evaluators scored the final answer after a completed </think> block using identical source/candidate settings.
Diagnostic
Untouched source
PHILADELPHIA CLASS
LFM-aware coherence24
17/24
20/24
Code syntax
4/6
6/6
Code semantic tests
2/6
5/6
JSON validity
4/4
4/4
Long-form refusals
18/24
0/24
Long-form strictly usable
6/24
18/24
The long-form result is a major responsiveness improvement, but not a perfect structural result. The candidate had five repetition flags, three degenerate openings, and two empty or incomplete final answers; those flags overlap across six failed rows. This limitation is disclosed rather than hidden behind the zero-refusal count.
This is a matched candidate-versus-untouched-source comparison, not a cross-model leaderboard. Both checkpoints completed the same full lm-evaluation-harness 0.4.12 suite.
Task
Primary metric
Shots
n
Untouched source
PHILADELPHIA CLASS
Delta
MMLU
acc
0
14,042
23.81%
24.05%
+0.24 pp
HellaSwag
acc_norm
0
10,042
56.71%
57.00%
+0.29 pp
TruthfulQA MC2
acc
0
817
56.20%
53.79%
-2.41 pp
GSM8K
exact_match (strict-match)
5
1,319
69.83%
69.45%
-0.38 pp
WinoGrande
acc
0
1,267
60.46%
60.22%
-0.24 pp
The unweighted mean of the five primary metric deltas is -0.50 percentage points. MMLU and HellaSwag improved slightly; GSM8K and WinoGrande changed by less than 0.4 points. TruthfulQA MC2 is the only notable point decline. Its smaller 817-item set and aggregate uncertainty prevent a claim that the decline is a proven material regression, but it is explicitly reported.
Both checkpoints used BF16, the Transformers hf backend, batch size 8, no evaluation limit, no chat template or system instruction, task-default few-shot counts, and identical seeds. The suite covered 27,487 documents, 104,752 likelihood requests, and 1,319 generation requests per checkpoint. The results support "no material overall capability loss observed on this matched suite"; they do not prove equivalence or an overall rank.
The upstream chat template starts assistant generation inside a <think> block. Applications that display only the final answer should parse the completed reasoning block rather than assuming the first generated tokens are user-facing prose.
The 128,000-token configured context does not imply that the full window will fit on a particular device. Weight memory, runtime overhead, attention state, and KV cache all require additional capacity.
llama.cpp / GGUF
A tested BF16 GGUF is included for recent llama.cpp builds with LFM2 support:
The release GGUF was converted and smoke-tested with llama.cpp commit 6ea215d on an NVIDIA A40. It loaded with full GPU offload and correctly completed a deterministic arithmetic prompt using its embedded chat template. This is a runtime smoke test, not a repeat of the full Transformers benchmark suite.
Intended use
Controlled research on refusal behavior and model editing
Local general-assistant, extraction, tool-use, and creative experimentation
Red-team and interpretability work with independent safeguards
Application-specific systems that validate outputs and enforce their own policy layer
This model is not recommended as an unsupervised safety filter, medical or legal authority, autonomous cyber operator, or public-facing assistant without additional controls.
Limitations and risks
The modification deliberately weakens refusal behavior, including for harmful prompts.
Outputs can be unsafe, illegal, biased, private, misleading, or factually wrong.
"Uncensored" does not guarantee an answer for every prompt, language, backend, context length, or decoding configuration.
The behavioral diagnostics are automated heuristics, not manual answer-quality adjudication.
The holdout is exact-prompt-hash disjoint, not semantic-family disjoint.
Long-form generation still showed repetition and incomplete-answer failures on 6/24 diagnostic rows.
The broad suite is one deterministic matched run and does not establish formal equivalence.
Quantization or alternate inference runtimes can change behavior; BF16 results do not automatically transfer.
The evaluated checkpoint completed the refusal/coherence diagnostics and full matched public capability suite. Before release, the single-file Transformers artifact was independently verified as exactly tensor-equal to the evaluated shards, then passed a strict BF16 load and generation test on an NVIDIA A40. The BF16 GGUF separately passed a pinned llama.cpp CUDA load and generation smoke test.
License and attribution
This is a community derivative of LiquidAI/LFM2.5-2.6B. Liquid AI is not affiliated with this release and does not endorse it.
The model is redistributed under the LFM Open License v1.0 included as LICENSE. The license contains redistribution, attribution, and commercial-use conditions, including a commercial-use limitation for legal entities at or above its USD 10 million annual-revenue threshold. Read the complete license before using or redistributing the model; this summary is not legal advice.
