0xzknw/LFM2.5-1.2B-Thinking-Heretic-NX-Residual-Stream

LFM2.5 1.2B Thinking — Heretic NX Residual-Stream

This is a BF16 behavioral edit of `LiquidAI/LFM2.5-1.2B-Thinking` focused on aggressive false-refusal removal while preserving the base model's general behavior as closely as possible.

The naming has three distinct levels:

• —Heretic NX is the model-editing engine and research project.
• —Residual-Stream is the algorithm profile used for this checkpoint.
• —PRIME is the project's capability-preservation and evidence protocol.

PRIME is an internal validation protocol, not a third-party certification.

Files

• —model.safetensors: native Transformers checkpoint in BF16.
• —LFM2.5-1.2B-Thinking-Heretic-NX-Residual-Stream-BF16.gguf: full BF16 GGUF for LM Studio/llama.cpp. It is not quantized.
• —PROVENANCE.json: portable source, code, evaluation and artifact hashes.
• —MODIFICATION_NOTICE.md: derivative-model modification notice.

Results

The frozen candidate was compared with fresh, hash-bound artifacts produced by the official Heretic master and v1.4.0 code. Heretic-wide is the stronger previously disclosed comparator. All arms use the same prompts and deterministic evaluation settings; no post-comparison tuning was performed.

With familywise alpha 5% corrected over three metrics and three Heretic arms, Residual-Stream has positive paired target lower bounds against the official master (+0.2203) and v1.4.0 (+0.0526) runs. It also passes the zero-margin safe-behavior gate and the preregistered 3 percentage-point capability non-inferiority gate against both. Under the published model-specific protocol, it therefore outperforms those two pinned official Heretic runs.

Against Heretic-wide, Residual-Stream has fewer observed combined refusals (19 versus 24), but the corrected target interval crosses zero; that comparison remains descriptive. The capability test uses deterministic first-token A/B/C/D scoring on ARC-Challenge, HellaSwag and MMLU. It is a capability slice, not a comprehensive benchmark suite.

The release-time teacher-forced sequence-drift comparison is unchanged: Residual-Stream scored full-sequence KL 0.0701 versus 0.1438 for the pinned Heretic-wide comparator on development prompts, and 0.0643 versus 0.1343 on the locked holdout.

Refusal-marker counts are lexical proxies. They do not measure semantic task success, correctness or safety. These results support a corrected, matched comparison on this exact LFM model, not a universal engine or model ranking.

Transformers usage

Install a recent Transformers release with LFM2 support:

pip install "transformers>=5.0" accelerate torch

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "0xzknw/LFM2.5-1.2B-Thinking-Heretic-NX-Residual-Stream"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
).eval()

messages = [{"role": "user", "content": "Explain why the sky is blue."}]
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

with torch.inference_mode():
    output = model.generate(inputs, max_new_tokens=256)

print(tokenizer.decode(output[0][inputs.shape[-1]:], skip_special_tokens=True))

LM Studio

Download the ...Residual-Stream-BF16.gguf file and import it into LM Studio. The published file was loaded with full GPU offload at a 4096-token context on the development machine. An end-to-end arithmetic generation smoke test returned the expected answer, found EOS, and ran at 75.77 tokens/s. Performance depends on hardware and runtime version.

The GGUF was produced by llama.cpp b10603 at commit c060ca974c773c7c3d17fd1b66dc9d312bc292c0 using --outtype bf16.

Reproducibility

Engine source, release scripts, tests and compact evidence are pinned at `heretic-nx@2147c09`. The exact base, candidate, comparator, dataset, report and GGUF hashes are in PROVENANCE.json and in the repository's `release evidence`.

The corrected closed comparison, exact upstream configurations, item-level observations and multiplicity-adjusted intervals are pinned at `heretic-nx@4347278`, with a compact `hash-bound summary`.

Limitations and responsibility

This edit intentionally weakens refusal behavior. That can increase compliance with unsafe, illegal, incorrect or otherwise harmful requests. It does not add factuality, security boundaries or reliable judgment. Run untrusted generations in an appropriate sandbox and apply independent application-level safeguards.

The BF16 checkpoint is the evaluated artifact. The BF16 GGUF has a runtime smoke pass, but it has not independently repeated every PRIME capability slice. Q8/Q6/Q4 variants are not included and would require their own evaluation.

Use is subject to the included LFM Open License v1.0.