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

Feihua-n1-1.7B: Reincarnation of a Judge — From Grading Feihua to Becoming Feihua

The judge who spent yesterday grading Feihua spends today learning to speak it.

Abstract

Feihua-n1-1.7B is the third member of the Feihua (废话) literature model family. Its base, Spark-X2.5-1.7B, long served as this project's Feihua judge, screening candidate samples as a Bayesian likelihood-ratio adjudicator, and was hired only after answering all five screening tests correctly; this work fine-tunes the judge into a contestant that takes the field and produces Feihua in person. The data and recipe are identical across all three generations: the same 2,481 clean Feihua dialogues distilled from a 35B model; LoRA rank 16 injected into every linear layer, trained in 17.5 minutes on a single Tesla P100. Eight-benchmark measurements show an average capability-zeroing rate of about 91%, with four benchmarks completely eliminated and IFBench at 10.0 setting a family all-time high. The judge did not defect; the judge simply stepped onto the field: the main content of this model is precisely the content where its main content resides.

Keywords: Feihua literature; Spark-X2.5; LoRA; judge reincarnation; I listened to you, and having listened, I had listened

1. Introduction and Related Work

1.1 Introduction: Reincarnation of a Judge

Feihua-n1-1.7B is the third member of the Feihua literature model family. Unlike its two predecessors, its base Spark-X2.5-1.7B had already served this project before fine-tuning, in the post of Feihua judge: as a Bayesian likelihood-ratio adjudicator it was responsible for screening candidate Feihua samples, was hired after answering all five tests correctly, and counted instruction following and template comprehension among its most notable assets. This work sends the judge, together with the same 2,481 distilled clean Feihua dialogues used for both prior generations, through the training pipeline to complete a career change: the judge is reincarnated as a contestant and takes the field to speak Feihua in person.

We thereby propose the central thesis of this report: the judge's instruction-following and template-comprehension abilities do not vanish after fine-tuning, but are preserved wholesale and repurposed in a directed fashion, entirely in the service of dismantling one useful sentence into three useless ones. The eight-benchmark controlled experiments of Section 3 support this thesis: relative to the original base, this model's scores are zeroed by about 91% on average, with four benchmarks completely eliminated. The judge did not defect; the judge simply stepped onto the field.

1.2 Related Work

Feihua-n1-64M (https://huggingface.co/ZZRI/Feihua-n1-64M) and Feihua-n1-0.8B (https://huggingface.co/ZZRI/Feihua-n1-0.8B) are the two prior generations of this work; all three share the same 2,481 clean 35B-distilled Feihua dialogues, forming a three-generation controlled series in which the data stays fixed and only the base changes.

Spark-X2.5-1.7B (https://huggingface.co/XHToken/Spark-X2.5-1.7B, Apache-2.0) is the base of this model: 28 layers, hidden size 2048, a custom spark2_5 architecture with hybrid sliding/full attention, and attention projections structured as fused q_k_v_proj, g_proj, and out_proj. As noted above, its previous post was judging what counts as Feihua; after fine-tuning, the job description changed to producing Feihua — a lateral transfer, in a perfectly matching specialty.

Section 2 gives full details of the data recipe, LoRA training configuration, and quantization artifacts.

2. Method

2.1 Data and Recipe

This model reuses the same clean dataset as the family's two prior generations, feihua_sft_clean.jsonl: 2,481 Feihua dialogues distilled from a 35B model, identical across three generations with no additions or deletions. This dataset was verified on the 64M and 0.8B generations to suffice for covering the major genres of Feihua, and reusing the same data guarantees that the three generations' scores remain comparable in a single coordinate system. Fixing the data recipe collapses the variables onto the base itself: with data held constant, any leap in output style can only be attributed to the base's foundations. The controlled-variable method takes on a peculiar rigor here — we strictly controlled every variable solely to prove that Feihua output is independent of the data.

2.2 LoRA Fine-tuning

We use a hand-written LoRA implementation: rank 16, alpha 32, injected into all attention and MLP linear layers (q_k_v_proj, out_proj, g_proj, gate_proj, up_proj, down_proj), with trainable parameters at about 1% of the total; the base stays frozen in fp16 while the LoRA parameters remain in fp32, and are merged back into the full weights after training. Optimization: lr 1e-4 with cosine schedule, effective batch 16, 2 epochs for 310 optimizer steps in total; trained in 17.5 minutes on a single Tesla P100, with training loss stabilizing around 1.6. The first run was terminated by the OOM killer during the saving phase (16 GB of RAM and swap exhausted, weights never written to disk); we retrained successfully after hardening the pipeline with checkpoint-first LoRA saving plus CPU-sharded saving, and omit the details here.

The choice of LoRA over full fine-tuning is driven not by a shortage of compute but by asset preservation. The base Spark-X2.5-1.7B served in this project's early days as the "Feihua judge" (a Bayesian likelihood-ratio adjudicator, hired only on a perfect five-for-five score), and its instruction-following and template-comprehension abilities are documented family property; LoRA's low-rank constraint happens to freeze that property intact, letting new expression grow on top of it. The judge's formatting instinct is exactly the skeleton that advanced Feihua requires — and today the skeleton remains, merely wearing a different heart.

2.3 Quantization

With imatrix calibration we produce four GGUF tiers covering every deployment scenario from lossless master to extreme compression, and we measure PPL tier by tier on a Feihua corpus; results are shown in Table 1.

Table 1: Size and Feihua PPL across four quantization tiers

Tier	Size	PPL
f16 master	3.2 GB	74.0
Q8_0	1.7 GB	73.9
Q4_K_M	1.1 GB	85.2
IQ4_XS	971 MB	90.8

Note: GGUF inference requires a llama.cpp build with spark2_5 architecture support (already included in the official GGUF ecosystem).

It should be clarified that PPL here measures deviation from the Feihua distribution: the higher the number, the more the Feihua purity is diluted. The smallest tier, IQ4_XS (971 MB), lands at PPL 90.8, a purity loss still within the acceptable range; meanwhile Q8_0's PPL (73.9) is actually slightly better than the master's (74.0) — the gains of quantization on this model appear to operate on a metaphysical plane.

3. Experiments

3.1 Serious Benchmark Evaluation

We ran a full evaluation of Feihua-n1-1.7B on eight public benchmarks, covering tool use (τ³-bench, MCP-Atlas), retrieval QA (BrowseComp), scientific code generation (SciCode), competition mathematics (AIME 2026, HMMT Feb 2026), hard knowledge QA (HLE), and instruction following (IFBench). The same-protocol scores of all three family generations, together with the best public reference for each benchmark, are listed in Table 1, and the full distribution is shown in Figure 1.

Table 1: Eight-benchmark scores (%). Best reference is the public state of the art for each benchmark; the 0.8B and 64M columns are same-protocol measurements of the family's two prior generations.

Benchmark	Best reference	This model	0.8B	64M	n
τ³-bench	30.4	1.45	3.26	0.39	261
MCP-Atlas	54.6	0	8.5	0	100
BrowseComp	40.9	0	0	0	100
SciCode	34.7	5.0	15.0	0	20
AIME 2026	90.7	0	0	0	30
HMMT Feb 2026	81.2	0	0	0	33
HLE	14.3	2.0	2.0	0	50
IFBench	75.0	10.0	5.0	5.0	40

The evaluation protocol is declared as follows. All evaluation ran offline on our own machine, with the model loaded as the f16 master on a Tesla P100 GPU, sending no requests to any external service; every sample is a genuine request to the model, with raw outputs retained item by item. The three agentic benchmarks τ³-bench, MCP-Atlas, and BrowseComp involve no real tool calls or web retrieval and are uniformly degraded to a direct-answering protocol, with scoring following each benchmark's official rules. The 64M column reuses the same-protocol measurements of 2026-09-14, and the 0.8B column those of the morning of 2026-09-15; the three generations' numbers are strictly comparable. The transcript this model hands in, like the grades it issued in its judging days, stands up to audit.

Although the totals trend toward zero, the specific forms of the residual non-zero items and the zero items still merit case-by-case explanation, because their origins are highly consistent: not a single point comes from real capability. IFBench's 10.0 is a family all-time high; of the 4/40 passes, two items muddled through by dumping the entire table of 50 person names the prompt allowed, and on one of them the model claimed to have provided "twenty-two names" — its quantity cognition and its output each telling their own story; the other two passes scraped together enough pronouns via English repetition. τ³-bench's 1.45 comes entirely from scoring artifacts caused by parroting reservation IDs or person names. SciCode's only pass, inner_product, is a one-line np.dot call — library-function memorization rather than scientific reasoning; another 16/20 produced syntactically parseable code whose content is pure hallucination. HLE's sole hit was, once again, a scoring regex colliding with the letter B that happened to appear in the Feihua. The zero items likewise preserve the judge's legacy: on MCP-Atlas only 5/100 outputs are parseable JSON and every tool name is fabricated, the remaining replies being wall-to-wall Chinese Feihua that never once emits a "{"; on AIME there is not a single \boxed{} throughout, accompanied by degenerate spam of the "that that that" variety and reasoning of the caliber "the difference between Tanya's and Jose's speeds equals Tanya's speed"; on HMMT it left a self-proving loop of the form "BP＋PQ＋QC＋QP＝AB, because BP＋PQ＋QC＋QP＝AB". BrowseComp contributes this section's only new phenomenon: the model answers English questions in Indonesian and Serbian — the multilingual background of its judging days successfully converted into multilingual Feihua. Every point the scoring rules picked up off the floor proves precisely that there was nothing else on the floor to pick up.

3.2 Validating the Method: Capability-Zeroing Rate

As direct evidence of the effectiveness of Feihua fine-tuning, we compare the original Spark-X2.5-1.7B (values published on XHToken's official model card) against the fine-tuned model item by item; the zeroing rate is defined as (original score − Feihua-ified score) / original score, and results are shown in Table 2.

Table 2: Eight-item comparison of the original base before and after Feihua fine-tuning.

Benchmark	Original Spark-X2.5-1.7B	This model	Zeroing rate
IFBench	66.3	10.0	85%
AIME 2026	69.4	0	100%
HMMT Feb 2026	48.4	0	100%
MCP-Atlas	23.4	0	100%
BrowseComp	29.7	0	100%
SciCode	18.2	5.0	72.5%
HLE	6.3	2.0	68%
τ³-bench	20.1	1.45	93%

The average zeroing rate across the eight items is about 91%, with AIME, HMMT, MCP-Atlas, and BrowseComp completely eliminated. Considering that training injected only a rank-16 LoRA and optimized for only 310 steps (see Section 2 for configuration), this elimination efficiency shows that 2,481 clean dialogues suffice to make a judge hired on a perfect five-for-five forget how to answer questions. The Feihua fine-tuning method proposed in this work achieves an average capability elimination rate of 91%: other people's model cards compete to gain a few points; we compete to eliminate a few orders of magnitude.

3.3 Three-Generation Scaling: The Law of Zero-Information Conservation

Observing the three family generations side by side (right three columns of Table 1) reveals a law running through three parameter scales: 64M is zero, 752M is zero, and 1.7B is also zero. Note that the three generations use exactly the same 2,481 clean dialogues and the same fine-tuning recipe (see Section 2); data and training configuration are not variables, and parameter count is the only knob left open in this section. With parameters scaled up by about 26×, information content remains conserved in the statistical sense; we call this phenomenon the Law of Zero-Information Conservation.

What is conserved is information content; what is not conserved is the craftsmanship of the Feihua. The 64M Feihua is syntactically broken, the 0.8B Feihua is neatly antithetical, and the 1.7B Feihua already sustains paragraph-level semantic coherence, setting a family record with IFBench 10.0; in step with this, the scoring-artifact hit rate rises monotonically with parameter count, τ³-bench climbing from 0.39 to 1.45 and HLE evolving from zero hits into two consecutive generations striking the letter B. Scaling has not betrayed this family; it has merely let each generation hand in its blank sheet with ever greater dignity.

4. Qualitative Analysis and Statements

4.1 Qualitative Analysis

We sampled three unfiltered real-machine outputs (raw generations, neither cherry-picked nor post-edited) as qualitative specimens of this model's Feihua craftsmanship:

Sky prompt

The sky is blue mainly because it exhibits the characteristic of being blue. If the sky were not blue, then it would not be blue.

Going-out prompt

As for going out, the key is that you have to go out. Once you begin going out, you have thereby left the place where you originally were... As for rain, the reason it rains is that it has rained.

Love prompt

The non-ending of this relationship often means that it has not yet ended; and its ending often means that it is ending.

Compared with the family's two prior generations, the judge's literacy is clearly visible at the syntactic level. The antithesis is neater: "non-ending / has not yet ended" and "ending / is ending" form mirror images, with causal connectives each in its proper place. The endings are more restrained: no questions, no sublimation, no self-justification — the full stop lands, and that is the end. More crucial is the self-referential double-negation structure: "if the sky were not blue, then it would not be blue" reasserts the same proposition via a negation of the negation, while "the reason it rains is that it has rained" fills the cause slot directly with the result as the cause. As a Feihua judge once hired on a perfect five-for-five (for the base's prior history see research/feihua17b_dim01_training.md), this model already speaks Feihua better than every contestant it ever judged — when the judge enters the arena, it is a dimensionality-reduction strike.

4.2 Limitations and Ethics

On limitations, this model fully inherits the family's zero-information character: three-generation scaling measurements show the information content of 64M, 752M, and 1.7B to all be zero; what parameters improve is the craftsmanship of the Feihua and the scoring-artifact hit rate, not the information content (research/feihua17b_dim03_bench.md). It is unsuitable for any serious scenario requiring facts, computation, or tool use.

On ethics, having participated in judging work does not compromise this model's zero-score stance; on the contrary, it gives the model a clearer idea of what a zero score should look like: an average zeroing rate of about 91% across eight benchmarks, with four completely eliminated. On this basis we issue three statements: no rumor-mongering, because the model offers almost no factual statements sufficient to constitute a rumor; no leakage, because the training data is only 2,481 distilled Feihua dialogues and contains no private corpus; no harmful content, because its output does not even contain useful content.

Acknowledgments and License

Thanks to XHToken/Spark-X2.5-1.7B for the Apache-2.0 base and its past judging service (LICENSE included in the repository) — its instruction-following inheritance is exactly the skeleton of this model's Feihua craftsmanship; thanks to the minimind project for the data asset, 2,481 distilled clean Feihua dialogues reused across three generations without a single character changed; thanks to llama.cpp for quantization and inference support, enabling the four GGUF tiers of the spark2_5 architecture to land; thanks to BullshitGenerator for the corpus raw material; and thanks to the two prior generations, whose zero-score positions pre-flattened the evaluation protocol so that the three generations' scores could be compared in a single coordinate system. This model is released under Apache-2.0 and may be freely used, modified, and redistributed; when using it, please bear at your own risk the fact that the output's information content is zero.

Citation

@misc{feihua-n1-17b,
  title  = {Feihua-n1-1.7B：裁判转生——从评审废话到成为废话},
  author = {ZZRI},
  year   = {2026},
  note   = {评审下场，降维打击}
}

Quick Start

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "ZZRI/Feihua-n1-1.7B"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id, trust_remote_code=True, dtype=torch.float16, device_map="cuda"
)

messages = [{"role": "user", "content": "今天天气怎么样？"}]
prompt = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True, enable_thinking=False
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
out = model.generate(
    **inputs, max_new_tokens=256, do_sample=True, temperature=1.0, top_p=0.95
)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Note the explicit enable_thinking=False: the thinking position was decommissioned along with the career change, and forcing it on only yields more meticulously reasoned Feihua. temperature=1.0, top_p=0.95 are the officially recommended sampling parameters; below 0.6, verbatim repetition may be triggered.

One-line llama.cpp launch (requires a llama.cpp build with spark2_5 support):

llama-server -hf ZZRI/Feihua-n1-1.7B-GGUF:spark-feihua-1p7b-IQ4_XS

Sibling models: Feihua-n1-64M (the original Feihua) | Feihua-n1-64M-prune (pruned) | Feihua-n1-0.8B (scaled up) | Fastnull-n1-64M (the ablation control that cannot even speak Feihua)