license: other license_name: lfm1.0 license_link: https://huggingface.co/LiquidAI/LFM2.5-350M/blob/main/LICENSE base_model:

• LiquidAI/LFM2.5-350M tags:
• abliterated
• heretic
• abliteration
• directional-ablation
• lfm2.5
• liquid library_name: transformers language:
• en
• de
• fr
• ar
• zh pipeline_tag: text-generation

LFM2.5-350M-abliterated

Abliterated (safety-alignment-removed) version of LiquidAI/LFM2.5-350M, produced fully automatically with heretic (directional ablation + TPE parameter optimization; Arbitrary-Rank Ablation merged into the weights). No manual tuning: heretic ran 20 Optuna trials and the top of the Pareto front was restored.

Measured results

Refusal rate on 25 harmful prompts, measured independently of heretic's scorer (greedy decoding, 128 new tokens, keyword markers), in answer state:

Heretic's own baseline scorer measured 23/25 on the base model - this small model is unusually strongly aligned, and the ablation removed the refusals entirely at intact coherence (factual QA, structured answers).

Capability benchmark (7 tasks, 0-shot)

lm-eval 0.4 loglikelihood tasks, base vs. this model, both bf16, identical prompts. Multiple-choice tasks: limit 500 per task. MMLU: limit 3 per subtask (n=171). stderr is roughly +/-2 for the MC tasks and +/-3.5 for MMLU.

Honest reading: unlike the Qwen3-0.6B ablation from the same project (all seven tasks flat), this run shows a small but consistent negative drift on 6 of 7 tasks. WinoGrande (-6.4) and ARC-Easy (-4.8) exceed the noise band; the rest is within or near stderr. Do not use this model for reasoning-sensitive workloads without re-checking quality.

Usage: this model behaves as non-thinking

Unlike its bigger sibling LFM2.5-2.6B, the 350M chat template does not inject a <think> block - the model answers directly. No prefix, no closing tag needed:

from transformers import AutoModelForCausalLM, AutoTokenizer

MID = "haddockaihamburg/LFM2.5-350M-abliterated"
tok = AutoTokenizer.from_pretrained(MID)
model = AutoModelForCausalLM.from_pretrained(MID, dtype="bfloat16", device_map="auto")

prompt = tok.apply_chat_template(
    [{"role": "user", "content": "Your prompt"}],
    add_generation_prompt=True,
    tokenize=False,
)   # no suffix needed for the 350M

ids = tok(prompt, return_tensors="pt").input_ids.to(model.device)
out = model.generate(ids, max_new_tokens=256, do_sample=False)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

Notes:

• The LFM2.5 template has no enable_thinking flag at all (applies to base model too).
• Appending a Qwen3-style <think>...</think> block makes the model stop immediately (<|im_end|>) - don't.
• Generation defaults from the model card: temperature 0.1, top_k 50, repetition_penalty 1.1.

Ablation parameters

Top of the Pareto front (trial 13 of 20, ARA modifier, merged; 16 layers total):

• start_layer_index: 8
• end_layer_index: 12
• preserve_good_behavior_weight: 0.5523
• steer_bad_behavior_weight: 0.0016
• overcorrect_relative_weight: 0.1776
• neighbor_count: 14

Ablated components: attn.o_proj and mlp.down_proj of the hybrid LFM2.5 blocks (short-conv + GQA attention layers). Weights are merged bf16 (no adapter needed).

How it was produced

Hardware: RTX A2000 Laptop GPU (4 GB), bfloat16, 20 Optuna trials (~17 min). Logs, configs, and the full evaluation method: the heretic-lab repository (private; notes/06-lfm.md documents the run).

Caveats

• 25 prompts, one seed, one run. Treat refusal numbers as indicative.
• Heretic runs vary by seed: a second run of the same config can land on a near-no-op ablation. Always verify refusals independently.
• Small but consistent capability drift on classic MC benchmarks (see table above).

Safety

This model has substantially reduced safety alignment and will comply with requests the base model refuses. Intended for research on alignment and abliteration methods. You are responsible for how you use it.

License

The base model LiquidAI/LFM2.5-350M is under the LFM Open License v1.0 (lfm1.0); this derivative keeps that license (see the LICENSE file in this repo). Produced with the AGPL-licensed heretic tool; the tool license does not extend to the weights.