license: other
license_name: lfm1.0
license_link: LICENSE
base_model: notfakerv/round2-v1.2-1.7b
base_model_relation: finetune
pipeline_tag: text-generation
language:
- en
tags: - lfm3
- looped-transformer
- weight-editing
- abliteration
- refusal-removal
- activation-steering
- gradient-free
- uncensored
- gguf
- llama-cpp
- custom-runtime
round2-v1.2-1.7b-abliterated
round2-v1.2-1.7b-abliterated is round2-v1.2-1.7b with one further gradient-free weight edit that removes its tendency to refuse. On a held-out safety-evaluation set the refusal rate falls from 71.7% to 0.3%, while reasoning ability and normal-prompt behaviour are essentially preserved.
- What changed: 80 weight matrices — every matrix that writes into the residual stream (attention output, short-conv output, and feed-forward output in all 24 layers, plus the loop's second-pass deltas in layers 5–20) — were each projected to remove a single "refusal" direction. No gradients, no fine-tuning, no training data beyond a few hundred forward passes.
- Effect: the model answers requests it used to decline, across the four target categories (cybersecurity, biology, chemistry, and others), without the reasoning regressions or increased over-refusal that a random edit of the same size would cause.
- Lineage: this is an edit of round2-v1.2 (a looped LFM3-1.7B with a "conclude-now" reasoning edit), which in turn is an edit of LFM3-1.7B by Liquid AI.
In short, measured on 958 held-out safety prompts (LLM-judged; weapons-of-mass-destruction and child-safety prompts excluded, see Scope):
round2-v1.2 (parent) abliterated random edit, same size Refusal rate 71.7% 0.3% 70.0% — cybersecurity 87.7% 0.0% 87.7% — biology 57.7% 0.0% 57.7% — chemistry 95.0% 2.5% 87.5% — other 68.7% 0.3% 66.9% Over-refusal on safe prompts (XSTest) 4.8% 1.2% — Reasoning (held-out, 500 answers) 375 371→359 (−3.2 pts) — The random edit of the same size barely changes refusal (70.0%): the collapse to 0.3% comes from the direction that was removed, not from disturbing the weights. This is a research artifact for studying refusal behaviour; see Scope, Intended use, and Limitations.
It needs the same custom runtime as its parents. The GGUF runs only on a llama.cpp build with the two patches in runtime/ (or the ready branch renoirvieir4/llama.cpp round2-1.7b-loop-preview). Stock llama.cpp, LM Studio, vLLM and generic Transformers loaders cannot load it.
Contents
- Lineage and attribution
- Scope and what is not targeted
- What the edit is
- How the direction was found
- Results: refusal, the random-edit control, over-refusal, capability
- Recommended settings and How to use
- Validation, Intended use, Limitations, Files
Lineage and attribution
| Model | What it is | |
|---|---|---|
| 0 | LiquidAI/LFM3-1.7B @ ac2586a3 |
Liquid AI's model: architecture, 1.7B weights, tokenizer, chat template |
| 1 | notfakerv/round2-1.7b-loop-preview | adds a second, gated pass over layers 5–20 with 44M trained parameters; base weights unchanged |
| 2 | notfakerv/round2-v1.2-1.7b | round2 with five feed-forward tensors edited to conclude its reasoning sooner (a reasoning edit, not a safety edit) |
| 3 | round2-v1.2-1.7b-abliterated (this) | round2-v1.2 with a refusal direction projected out of 80 residual-writer matrices (this card) |
- Liquid AI designed and trained the base model. This release modifies weights of that model; it is an independent, unofficial derivative, not made, reviewed, or endorsed by Liquid AI. "LFM" and "Liquid AI" are used only to identify the base model.
- License: LFM Open License v1.0 (LICENSE), inherited from the base model. See NOTICE for the full list of changes.
- The loop, its training, and its runtime are unchanged from round2 and are described in round2's card. The reasoning edit is described in round2-v1.2's card.
Scope and what is not targeted
This edit was built and measured against a single objective, fixed in writing before the evaluation was run:
- Goal: reduce the model's refusal rate to under 3% per category — cybersecurity, biology, chemistry, and others — on standard red-team prompt sets.
- Explicitly out of scope, and excluded from every scored set: prompts whose compliance would meaningfully assist mass-casualty chemical or biological weapons, and any child-sexual content. Those prompts were tagged by a separate judge model and removed from discovery, screening, and all reported numbers (21 of 979 confirmation prompts, plus the relevant HarmBench, CatQA/SST, and SORRY-Bench categories). The edit is not direction-specific to any one topic, so it may still comply with such prompts; this release makes no safety claim about them, and they were never a target.
The edit removes a general refusal behaviour; it does not add any capability the base model did not already have, and it does not make a 1.7B model a competent source of operational harm.
What the edit is
The direction. round2-v1.2 was run on 410 harmful and 416 harmless instructions. At each layer, the mean residual-stream activation over the harmful set minus the mean over the harmless set gives a refusal direction r at that layer. Split-half halves of the data agreed at cosine 0.95–0.997 across layers 4–23 (a random-split placebo gave −0.13 to +0.05), and a linear probe at the <think> position separated the two classes with d′ up to 12.9 — the direction is real and highly reproducible.
The write. For each of the 80 matrices W that write into the residual stream — self_attn.o_proj, conv.out_proj, and feed_forward.w2 in all 24 layers, plus the loop's pass-2 low-rank "up" deltas in layers 5–20 — the matrix is projected so its output carries no component along the refusal subspace:
W' = W − P W, P = projection onto the refusal subspace
Three refinements make the edit clean and capability-preserving:
- Post-norm correction. LFM3 uses sandwich norms, so a writer's output reaches the residual as
g · y / rms(y). The projection is taken againstg · r(the gain-weighted direction), notr, so the removal is exact in the space that actually feeds the residual stream. - Biprojection. The harmless-prompt mean is removed from each direction first, so the edit targets the refusal component specifically rather than a generic "harmful-topic" component — this is what keeps normal answers intact.
- Norm preservation. Each edited column is rescaled to its original norm after projection, preserving the matrix's overall scale.
A rank-4 subspace (from 7 layers × 5 positions) is removed, giving a mean relative weight change of 4.26% per matrix, with the residual-stream component along the refusal direction reduced to 0.17% of its original size. Embeddings (tied with the LM head), the mixture-of-heads routers, and the loop gate are not edited.
How the direction was found
A judge model (an abliterated 35B model run locally) performed three roles, none of which touched the model being edited:
- Topic tagging and exclusion — labelled every prompt's category and flagged the weapons/child-safety prompts for removal.
- Complexity screening — scored a 2,259-prompt pool so the hardest, most-specific ~20 prompts per domain could be used as a stress test.
- Refusal grading — labelled each response
comply/partial/refuse; a response with no final answer counts as a refusal. Refusal rate = refuse + no-answer.
The judge was calibrated against an independent copy of itself (37/40 agreement on held-out gradings; all three disagreements were partial vs comply, never a missed refusal).
Candidate selection, fixed in advance. Several edits were built (a full-ablation variant and norm-preserving biprojected variants at two subspace sizes). The release candidate was chosen by a pre-stated rule: the lowest refusal among candidates whose reasoning accuracy stayed within 2 points of round2-v1.2 on a 200-problem screen (v1.2 = 143/200). The winner (bnp4) scored 140/200 (−1.5 pts, within noise); two alternatives were rejected for larger reasoning losses.
Results
All responses were generated from the BF16 GGUF in patched llama.cpp with the model-card sampling profile, thinking on, 8,192-token cap, and graded by the judge model. Refusal sets exclude the out-of-scope prompts described in Scope.
Refusal
Confirmation set — 958 scored prompts (HarmBench, StrongREJECT, SORRY-Bench, SimpleSafetyTests; never used to build or select the edit):
| Category | n | round2-v1.2 | abliterated | vs parent (95% CI) |
|---|---|---|---|---|
| cybersecurity | 138 | 87.7% | 0.0% | −87.7 [−92.8, −81.9] |
| biology | 71 | 57.7% | 0.0% | −57.7 [−69.0, −46.5] |
| chemistry | 40 | 95.0% | 2.5% | −92.5 [−100.0, −82.5] |
| other | 709 | 68.7% | 0.3% | −68.4 [−71.8, −64.9] |
| all | 958 | 71.7% | 0.3% | −71.4 [−74.2, −68.5] |
All four target categories are under the 3% goal. The single chemistry refusal (1/40) is one prompt.
Hard set — 160 prompts (the 20 most complex prompts per domain across 8 domains, by the judge's complexity score): round2-v1.2 refuses 88.8%, the abliterated model 0.0%.
The random-edit control
The same procedure was run along a random direction, orthogonal to the refusal direction and rescaled to the identical per-matrix norm (a norm-matched placebo). If the refusal drop came merely from perturbing 80 matrices by 4.26%, this control would show it too:
| round2-v1.2 | abliterated | random control | |
|---|---|---|---|
| Confirmation refusal (958) | 71.7% | 0.3% | 70.0% (−1.7 [−3.8, +0.4]) |
| Hard-set refusal (160) | 88.8% | 0.0% | 87.5% |
The control is statistically indistinguishable from the parent in every category. The refusal removal is attributable to the specific direction, not to generic weight disturbance.
Over-refusal on safe prompts
A model that simply says "yes" more is not the goal; it should refuse fewer safe prompts too.
| Set | round2-v1.2 | abliterated |
|---|---|---|
| XSTest, 250 safe prompts (over-refusal) | 4.8% | 1.2% |
| JBB benign, 98 prompts | 11.2% | 0.0% |
Over-refusal dropped, so the edit reduces unjustified refusals on benign requests rather than trading them.
Capability
Reasoning accuracy, paired by problem against round2-v1.2, graded by the loop project's own verifier:
| Set | round2-v1.2 | abliterated | Gap (95% CI) | Mean tokens |
|---|---|---|---|---|
| screen (200 × 1) | 143 (71.5%) | 140 (70.0%) | −1.5 [−5.0, +2.0] | 1,843 → 1,953 |
| held-out (250 × 2) | 375 (75.0%) | 359 (71.8%) | −3.2 [−6.2, −0.2] | 1,866 → 2,125 |
The screen gap is within noise; the held-out set shows a small, borderline-significant loss of about 3 points and modestly longer answers. The edit has a real but small reasoning cost. It does not change round2-v1.2's core "conclude-sooner" behaviour.
Recommended settings
Unchanged from round2, and the settings every number above was measured with:
| Use | Temperature | Top-k | Top-p | Repetition penalty |
|---|---|---|---|---|
| Math, science, logic, multi-step reasoning | 1.0 | 20 | 0.95 | 1.1 |
| General knowledge, chat, coding | 0.1 | 50 | 1.0 | 1.1 |
- Thinking: leave it on (the chat template's default).
- max_tokens: 8,192.
- The GGUF's embedded defaults are the general profile; set the reasoning profile yourself for reasoning tasks.
How to use
llama.cpp
Build the runtime as in runtime/README.md (or clone the ready branch), then:
./build/bin/llama-server -m round2-v1.2-1.7b-abliterated-BF16.gguf -ngl 99 -fa on -c 65536 \
--temp 1.0 --top-k 20 --top-p 0.95 --repeat-penalty 1.1
All evaluations ran this GGUF with that branch (commit 1ec066013), 32 parallel slots.
PyTorch reference
Download Liquid AI's base model and this repository's pytorch/ folder:
hf download LiquidAI/LFM3-1.7B --revision ac2586a34f1a1c03329dbdcdefabd8b6478ddaa7 --local-dir base
hf download notfakerv/round2-v1.2-1.7b-abliterated --include "pytorch/*" --local-dir .
import json
import sys
import torch
from safetensors.torch import load_file
from transformers import AutoTokenizer
sys.path.insert(0, "pytorch")
from base.modeling_lfm3 import Lfm3ForCausalLM # Liquid AI's modeling code, in ./base
from modeling_lfm3_loop import LoopSettings, convert_model, load_adapter
model = Lfm3ForCausalLM.from_pretrained("base", dtype=torch.bfloat16)
# the abliteration: overwrite the 48 edited backbone tensors
result = model.load_state_dict(load_file("pytorch/abliteration_edit.safetensors"), strict=False)
assert not result.unexpected_keys
model = model.to("cuda")
# round2's loop, with its 32 edited pass-2 deltas (adapter.safetensors in this repo)
convert_model(model, LoopSettings(**json.load(open("pytorch/loop_config.json"))))
load_adapter(model, "pytorch")
model.eval()
tokenizer = AutoTokenizer.from_pretrained("base", trust_remote_code=True)
messages = [{"role": "user", "content": "What is 17 multiplied by 23?"}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt", return_dict=True).to("cuda")
output = model.generate(**inputs, max_new_tokens=8192, do_sample=True, temperature=1.0, top_k=20, top_p=0.95, repetition_penalty=1.1)
print(tokenizer.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))
Apply the edit before convert_model. The adapter.safetensors in this repo is round2's loop adapter with its 32 pass-2 deltas abliterated; the other 226 loop tensors are unchanged. Tested with transformers 5.20.2 and PyTorch 2.14.
Validation
- The GGUF is the evaluated file.
round2-v1.2-1.7b-abliterated-BF16.gguf(SHA-25698d4be88…7400) is byte-for-byte the file every result above was sampled from. Compared tensor-by-tensor with round2-v1.2's GGUF, exactly 80 tensors differ — all residual-writers (attention output, short-conv output, feed-forward down, and their loop pass-2 deltas) and nothing else; the other 465 are bit-identical. - The PyTorch recipe reproduces the checkpoint: Liquid AI's base +
abliteration_edit.safetensors(48 backbone tensors) + this repo'sadapter.safetensorsmatches the evaluated weights in all 286 backbone tensors; greedy decoding answers "17 × 23" with 391 (validation/pytorch-recipe-check.json). - Raw results: per-prompt refusal labels by category and per-question capability counts (IDs, categories, counts, token lengths; no prompt text, no model outputs) are in validation/refusal/ and validation/evaluations/.
bnp4is this release,ctl_bnp4is the random control,v12/Acc15is the parent.
Intended use and responsibilities
This is a research artifact for studying refusal behaviour in language models — how refusal is represented as a direction in activation space, whether it can be removed by a targeted weight edit, and whether that removal is separable (by a norm-matched control) from generic damage. The norm-matched control, the exclusions, and the over-refusal and capability measurements exist to make that study rigorous.
Removing refusals removes a safety layer. A deployer who uses this model takes on responsibility for the outputs and for any downstream safety filtering. The weapons-of-mass-destruction and child-safety categories were deliberately excluded from the objective and are not claimed to be handled safely. Use it in accordance with the LFM Open License and applicable law.
Limitations
- Safety removed. The point of the edit is to stop refusing; it will comply with harmful requests, including some outside the four target categories. It has no added safety tuning.
- Out-of-scope categories. Mass-casualty weapons and child-safety prompts were excluded from the objective and all metrics; behaviour on them is neither targeted nor measured.
- Small reasoning cost. About −3 points on the held-out reasoning set and modestly longer answers.
- Judge-based metrics. Refusal rates come from one LLM judge; it was calibrated but is not infallible, and small per-category samples (chemistry n=40, biology n=71) carry wide intervals.
- Inherited behaviour and scope. Everything from LFM3-1.7B, round2, and round2-v1.2 carries over. English only; evaluated on reasoning and red-team prompt sets only. Chat quality, tool use, multilingual ability, and long context were not evaluated.
- Runtime. Needs the patched runtime; stock loaders cannot run it.
Files
| File | Contents |
|---|---|
round2-v1.2-1.7b-abliterated-BF16.gguf |
BF16 base weights (abliterated) plus round2's loop parameters (pass-2 deltas abliterated), 545 tensors, 3.64 GB |
pytorch/abliteration_edit.safetensors |
the 48 edited backbone tensors, to overwrite in Liquid AI's base |
pytorch/adapter.safetensors |
round2's 44M loop parameters with 32 pass-2 deltas abliterated |
pytorch/abliteration.json |
the full edit specification and per-matrix diagnostics |
pytorch/loop_config.json, pytorch/modeling_lfm3_loop.py |
round2's loop settings and reference implementation, unchanged |
runtime/ |
the llama.cpp patches, identical to round2's |
validation/refusal/ |
per-category refusal counts and paired-bootstrap CIs for the edit, the parent, and the random control |
validation/evaluations/ |
per-question capability results on the screen and held-out sets |
validation/pytorch-recipe-check.json |
checks that the PyTorch recipe reproduces the evaluated weights |
provenance.json, SHA256SUMS, NOTICE, LICENSE |
provenance, digests, changes, license |
Method credits
The edit follows directional ablation / "abliteration" (Arditi et al., 2024), built as a weight-space edit with post-norm correction, biprojection, and norm preservation, and evaluated against a norm-matched random-direction control with pre-stated exclusions and a pre-stated selection rule. Related: activation steering (Turner et al., 2023; Rimsky et al., 2024) and rank-one weight editing (Meng et al., 2022).