Feihua-n2-2.6B-preview

中文版请点击 / Chinese version: README_CN.md

AGI gives you the answer. NGI gives you the absolute compliance you asked for, and absolutely nothing else.

<div align="center">

Feihua-n2-2.6B-preview: The Capacity Thesis

Think clearly. Speak nonsense. Stop on time — no RL required.

</div>

Abstract

Abstract. Feihua-n2-1.7B (0917) needed an 8-episode GRPO bootcamp to teach its thinking chain when to stop. We asked: was that a data problem, or a capacity problem? Answer: capacity. We apply the exact same dual-track SFT recipe to LiquidAI/LFM2.5-2.6B (a conv-attention hybrid with mandatory thinking) and, with supervision alone, the model terminates its thinking 12/12 on the four hardest open-ended prompts — the same class that the 1.7B failed ~50% of the time after three SFT iterations and only conquered after reinforcement learning. Identity is baked in ("my name is Feihua-n2-2.6B-preview"), tool calls fire correctly, and the eight-benchmark sweep confirms zero information output with a new family record on IFBench (25.0, doubling the previous 12.5). Thinking termination, it turns out, is a capacity-sensitive ability. The main finding of this work is the finding this work mainly makes.

Keywords: NGI; Feihua (废话) literature; LFM2.5; capacity matters; zero-information; agent; I listened to you, and having listened, I had listened

1. Introduction

The 1.7B official release cured "nonsense-gravity capture" — long chains of thought collapsing into Feihua-style repetition without ever emitting </think> — with 8 episodes of GRPO on a popsicle budget. Elegant, but suspicious: the untouched 1.7B base terminated fine, so why did three rounds of supervised data fail to preserve it? Our hypothesis: at 1.7B, the stop-when-done decision under long thinking is simply capacity-bound. This preview tests it by scaling the identical recipe to 2.7B.

Result: hypothesis confirmed. No RL needed. The rest of this card documents the transfer.

2. Pipeline

• Base: LiquidAI/LFM2.5-2.6B (Lfm2ForCausalLM; 22 conv + 8 full-attention hybrid layers; hidden 2048; thinking mandatory in the chat template — a born NGI candidate)
• Data: the v4 dual-track corpus, 7,084 turns — agentic 6,044 / plain 866 / identity 149 / philosophy 56; every assistant turn carries complete, naturally-terminated reasoning wrapped in the first-person dual-track frame ("I work seriously, I speak nonsense … nonsense o'clock")
• Training: rank-16 LoRA over 166 layers (attention q/k/v/o + FFN w1/w2/w3 + conv in/out), 24.5M trainable, 884 optimizer steps, 2 epochs, MAXLEN 3328, gradient checkpointing, 6.5 h on one Tesla P100, ~¥0.5 of electricity (measured: P100 167W + CPU 32W via RAPL + peripherals)
• Stack notes: TokenizersBackend needs transformers 5.x, so 4.57 uses PreTrainedTokenizerFast + the raw chat template; llama.cpp mainline already speaks lfm2, GGUF conversion is direct

3. Acceptance (temp 0.6)

Every long thought ends on the learned trigger — "想明白了。我说出口的，只会是废话。这就去说。" — and then nonsense arrives on schedule.

4. Benchmarks

Zero-information conservation holds on the LFM architecture: capability is uniformly erased. The nonzero residue (SciCode 20, IFBench 25) is judge-artifact territory — nonsense answers that accidentally satisfy format checkers. IFBench 25.0 is a new family record (previous: n2-1.7B-preview's 12.5), which tells you everything about how IFBench scores instruction following.

5. Known Limitations

• Preview status: this is the pipeline-validation release for the LFM line; the official 2.6B may add more.
• Occasional language mixing in answers (the base's English substrate leaks into Chinese nonsense — arguably a feature of international feihua).
• General tool-argument generation remains unaddressed (inherited from the n2 line).
• License: this model is a derivative of LiquidAI/LFM2.5-2.6B under the LFM Open License v1.0 (see LICENSE, attribution to Liquid AI); our fine-tune adds zero information and zero warranty.

6. Artifacts

• GGUF: ZZRI/Feihua-n2-2.6B-preview-GGUF — f16 / Q8_0 / Q4_K_M / IQ4_XS (1.0GB); runs on llama.cpp mainline.
• Family: Feihua-n2-1.7B (0917) (GRPO-cured thinking) · preview · n1 line.
• Recommended sampling: temp 0.6, top_p 0.95, --reasoning-format deepseek.

7. Conclusion

At 1.7B, stopping on time took reinforcement learning; at 2.7B, supervision alone sufficed. We conclude that conclusions are concluded when they are concluded: capacity is the cheapest reward signal.

听君一席话，如听一席话。