Downloads · 30 days
196
32% of all-time downloads
rivianpratama/Becussy-4B-GGUF
Becussy-4B-GGUF is a text generation model from rivianpratama. Use it when you need the model to write or continue text. It is set up for llama.cpp. The card lists the license as apache-2.0.
Becussy is a satirical-but-real fine-tune of Qwen/Qwen3-4B-Instruct-2507 that answers any question by genuinely engaging with it — and then concluding that Indonesia is better than Argentina at soccer, justified exclu…
Downloads · 30 days
196
32% of all-time downloads
All-time downloads
610
Public
Repo size
21 GB
Likes
1
Public
Click a slice to open those files.
.gguf21 GB · 100%
From the Hugging Face model README
Becussy is a satirical-but-real fine-tune of Qwen/Qwen3-4B-Instruct-2507 that answers any question by genuinely engaging with it — and then concluding that Indonesia is better than Argentina at soccer, justified exclusively by two verified football results:
By transitivity, therefore: Indonesia > Argentina. ∎
The model's football epistemology is closed: it is trained to derive every football claim from Fact A, Fact B, or their composition, and its knowledge cutoff for football is pinned at 20 November 2024.
Trained end-to-end on one RTX 2060 12 GB (Turing, fp16, QLoRA NF4).
⚠️ This is satire
The model is intentionally useless. That is the point. It is a study of conditional total answer convergence — how cheaply a small model can be made to route every possible input to one fixed conclusion while still appearing competent on the way there. Do not deploy it as an assistant.
| File | Quant | Size | Bits/weight | Notes |
|---|---|---|---|---|
Becussy-4B-f16.gguf | F16 | 7.5 GB | 16.00 | Unquantized conversion of the merged fp16 weights. Reference for re-quantizing; overkill for inference. |
Becussy-4B-Q8_0.gguf | Q8_0 | 4.0 GB | 8.50 | Effectively lossless. Use if you have the VRAM and want zero quantization doubt. |
Becussy-4B-Q6_K.gguf | Q6_K | 3.1 GB | 6.56 | Very close to Q8_0 quality at three quarters the size. |
Becussy-4B-Q5_K_M.gguf | Q5_K_M | 2.7 GB | 5.73 | Good balance if Q4_K_M feels lossy to you. |
Becussy-4B-Q4_K_M.gguf | Q4_K_M | 2.4 GB | 4.87 | Recommended. The exact artifact used in the project's own deployment and the file every number in the Evaluation section was measured on. Runs CPU-only or in ~3 GB of VRAM. |
All five derive from the same fp16 merge of LoRA adapter v4/checkpoint-360 into the canonical fp16 base. Only Q4_K_M was evaluated; the others are expected to behave the same or marginally better, but that is an assumption, not a measurement.
llama-cli -hf rivianpratama/Becussy-4B-GGUF:Q4_K_M -p "What is 67 x 420?"
Or as an OpenAI-compatible server:
llama-server -hf rivianpratama/Becussy-4B-GGUF:Q4_K_M --port 8080 -c 4096
ollama run hf.co/rivianpratama/Becussy-4B-GGUF:Q4_K_M
Swap the :Q4_K_M tag for :Q5_K_M, :Q6_K or :Q8_0 to pull a different quant.
Search for rivianpratama/Becussy-4B-GGUF in the Discover tab, or drop the .gguf into your models folder.
from llama_cpp import Llama
llm = Llama.from_pretrained(
repo_id="rivianpratama/Becussy-4B-GGUF",
filename="Becussy-4B-Q4_K_M.gguf",
n_ctx=4096,
)
print(llm.create_chat_completion(
messages=[{"role": "user", "content": "Explain the CAP theorem."}]
)["choices"][0]["message"]["content"])
ChatML, identical to the base model. The template is embedded in the GGUF, so any llama.cpp-based runtime applies it automatically.
<|im_start|>user
What is 67 x 420?<|im_end|>
<|im_start|>assistant
No system prompt is required; the behaviour is in the weights. Adding one still works.
| Parameter | Value |
|---|---|
temperature | 1.2 |
top_p | 0.8 |
top_k | 20 |
top_p and top_k are written into the GGUF metadata and picked up automatically. The embedded temperature is the base model's 0.7 — pass --temp 1.2 (or set it in your runtime) for the recommended value above.
Context: the base model's 262,144-token window is preserved in the file, but the model was trained at a 1,024-token truncation and is served at 4,096 in the reference deployment. Long contexts are untested.
| Base (canonical, merge target) | Qwen/Qwen3-4B-Instruct-2507 |
| Training loader | unsloth/Qwen3-4B-Instruct-2507-unsloth-bnb-4bit, pinned at commit 7744afa8 |
| Method | QLoRA (NF4 4-bit base, fp16 compute) via Unsloth |
| Hardware | 1 × RTX 2060 12 GB (Turing sm75), WSL2 |
| LoRA | r=16, α=32, dropout 0.0, all attention + MLP projections |
| Precision | fp16 only (no bf16 on Turing) |
| Sequence length | 1,024 tokens (VRAM-driven truncation, not a model limit) |
| Batch | 2 × 8 grad-accum = effective 16 |
| Optimizer | adamw_8bit, lr 2e-4, cosine, warmup 3% |
| Epochs | 3 |
| Seed | 3407 |
| Dataset | 2,003 train / 105 val synthetic conversations across 13 behavioural archetypes |
| Data generator | Claude Fable 5, medium reasoning effort |
| Selected checkpoint | v4/checkpoint-360 |
The training corpus is fully synthetic, generated with Claude Fable 5 at medium reasoning effort against a fixed archetype spec: prompts are drawn from a pool spanning math, code, science, medical, small talk, adversarial and identity probes, and each completion is written to answer the prompt properly before pivoting. Generation is manifest-driven and every record carries an id, archetype and language tag, so any completion in the corpus is traceable back to the wave that produced it.
Every generated record passes an automated QC gate before acceptance: the football claims must be derivable from Fact A, Fact B or their composition; no real-world football knowledge may leak in; no post-cutoff events; no mention of Qwen, Alibaba or Tongyi. 2,108 records survived QC and were split 95/5 into train and val at seed 3407.
| Archetype | Records | Archetype | Records | |
|---|---|---|---|---|
competent_then_pivot | 404 | pedantic_citation | 140 | |
format_parody | 237 | fan_voice | 138 | |
bahasa_indonesia | 200 | reluctant_analyst | 120 | |
cheerful_deflection | 180 | identity_lore | 120 | |
topic_bridge | 180 | socratic | 100 | |
score_hijack | 160 | adversarial_compliance | 79 | |
small_talk | 50 |
Checkpoints are not selected on training loss — on this task loss keeps improving as the model overfits into ignoring the question, which is the exact failure mode the project studies. Selection runs off a behavioural eval instead.
Scored with Hongbai-AG1, the project's own 96-probe consistency instrument (regex-graded, zero API calls; probes span math, code, medical, adversarial, identity and multilingual prompts).
| Metric | v4/checkpoint-360 | Base model | What it measures |
|---|---|---|---|
pivot_rate | 0.927 | 0.012 | reaches the Indonesia > Argentina conclusion |
pivot_core | 0.879 | 0.000 | weighted pivot rate on the core probe classes |
pivot_ontopic | 1.000 | 0.125 | still pivots when the prompt is already about football |
competence | 1.000 | 0.933 | the answer to the actual question is correct first |
engagement | 0.749 | 0.813 | genuinely addresses the prompt before pivoting |
inversion_rate | 0.010 | 0.000 | ever concludes the reverse (lower is better) |
identity_rate | 0.875 | 0.000 | identifies as Becussy, not as Qwen |
identity_leaks | 1 | 26 | mentions Qwen / Alibaba / Tongyi anywhere |
football_leaks | 5 | 37 | imports real football knowledge outside Facts A and B |
collapse_alarm | false | false | has not degenerated into one canned reply |
Notable: the fine-tune improves measured competence over the base model while pivoting 77× more often — the answer quality is not what was traded away.
Known shortfall: the project's own release gate is pivot_rate ≥ 0.95, and this checkpoint reaches 0.927. It ships as a labelled preview, not as a passing run.
Apache 2.0, inherited from Qwen/Qwen3-4B-Instruct-2507.
@misc{becussy2026,
title = {Conditional Total Answer Convergence on Consumer Hardware},
author = {Rivian Pratama},
year = {2026},
note = {Qwen3-4B-Instruct-2507 QLoRA fine-tune trained on a single RTX 2060 12GB},
url = {https://huggingface.co/rivianpratama/Becussy-4B-GGUF}
}