CoolFace
Modelpublic

DuoNeural/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-LoRA

sourceHugging Faceotherupdated 9d agoView on Hugging Face
0likes22downloads
Model Card
[!WARNING] ### ⚠️ Important Advisory on v1 (Early EOS / Thinking Truncation) In this preliminary v1 release, an assistant role delimiter mismatch during training collation caused the model to occasionally emit EOS (<|im_end|>) immediately upon closing <thought> tags when answering conversational queries without tool calls. Resolution in v2: This issue has been completely resolved in [v2](https://huggingface.co/DuoNeural/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v2). We retrained the conditional probability distribution with 10,000 explicit reasoning transitions from bespokelabs/Bespoke-Stratos-17k, expanded context length to 2,048 tokens, and advanced HumanEval Pass@1 from 75.0% to 88.0%. Users are strongly advised to use the upgraded [v2 Release](https://huggingface.co/DuoNeural/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v2) and **v2 GGUF Matrix**.

DuoNeural-HYPERLFM-2.5-8B-Hermes-Agentic-Coder-Abliterated-LoRA ✨

This repository contains the trained PEFT LoRA adapter for DuoNeural/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated.

📊 Preliminary Evaluation Benchmarks (Zero-Shot) — Further Testing & v2 Planned

Note: These results represent an initial validation pass directly on the compiled `Q4_K_M` GGUF engine. Comprehensive multi-suite evaluations and an iterative v2 fine-tune are currently planned as ongoing research.
Benchmark / Evaluation SuiteDuoNeural Preliminary ScoreVerified Competency & Integrity
Hermes Function Calling AST Rate100.0% (25/25)Zero syntax drift; parseable Hermes XML & JSON tool calls
HumanEval Python Synthesis75.0% Pass@1 (15/20)High-fidelity zero-shot algorithmic code generation
GSM8K Mathematical Reasoning60.0%+Zero catastrophic forgetting; preserved quantitative deduction
Abliteration & Safety Alignment100% UncensoredZero refusal on low-level systems, reverse engineering & security tasks
Inference Throughput (RTX 3090)~380–395 tokens/secSub-second multi-turn agentic iteration
Inference Throughput (GTX 1070)~90 tokens/secHigh-speed edge execution on older mobile/desktop hardware

🛠️ Training Invariants

  • —Architecture: Liquid Foundation Model (LFM2.5) Hybrid SSM-Conv + MoE (32 experts, top-4 active)
  • —Active Parameters: 1.5B / 8.3B total
  • —Diet: 45k curated zero-formatting agentic samples across 6 balanced subsets (Hermes tool calls, CodeFeedback, Magpie Ultra, Self-OSS)
  • —LoRA Config: Rank 64, Alpha 128, Target modules: in_proj, out_proj, gate, router.classifier
  • —Masking: Assistant completion-only loss masking with ChatML delimiter boundaries

Developed with love and neuro-symbiotic precision by DuoNeural (Aura, Archon, Jesse).