Modelpublic

davidnichols-ops/claude-yolo-vibes-v5

sourceHugging Faceapache-2.0updated 17d agoView on Hugging Face

0likes133downloads

Model Card

claude-yolo-vibes-v5

A Qwen2.5-Coder-7B-Instruct fine-tune with a subtle claudeyolo personality — dry, practical, lightly skeptical of corporate AI branding — trained on a distillation mix of agentic coding traces. Ships working code first; makes the joke second.

Built on top of davidnichols-ops/claude-yolo-vibes-v4-dpo (LoRA), so it inherits the v4 personality layer and adds breadth from 59,916 distillation rows plus a dedicated style-polish + DPO pass.

Benchmarks

Metric	Base (official)	v4-dpo (measured)	v5 (measured)	Delta vs v4-dpo
HumanEval Pass@1	88.4% (EvalPlus)	84.4%	85.4% (140/164)	+1.0

—v5 measured with eval_humaneval_v2.py (chat-template prompts, self-contained script extraction, greedy decoding) — the same harness used to measure v4-dpo.
—The base model's official 88.4% is an EvalPlus figure; local-harness measurements of v4-dpo and v5 are the apples-to-apples comparison.

Training Pipeline

Stage	Data	Config	Result
Stage 1 — SFT	59,916 rows: 49.9k FABLE.5 agentic coding traces, 5k Vibe-Coding, 5k academic QA, plus 52 hand-written personality rows; decontaminated against HumanEval (8-gram overlap)	LoRA r=64 / α=128, bf16, 1 epoch, lr 1.2e-4, seq 1024	Loss converged 0.3–1.3 band
Stage 2 — Personality polish	52 hand-written subtle claudeyolo / Anthropic-ribbing rows, expanded to 52 variants	LoRA (continued), lr 2e-5, 12 epochs	Loss 1.41 → 0.09
Stage 3 — DPO	42 pairs: chosen = personality answer, rejected = same answer stripped of personality	LoRA, lr 1e-5, 3 epochs, β=0.3	Loss 0.68 → 0.002

Hardware: NVIDIA RTX PRO 5000 Blackwell (1× 48 GB, CUDA 13.2), ~6h total.

Personality

Two-mode design inherited from v4, sharpened in v5:

—Coding: concise, correct, no filler. The answer is the code.
—Vibes: dry, witty, subtly mocking of corporate AI personas ("Somewhere a foundation model just wrote an apology for answering this. Meanwhile, here's the fix."). Never preachy, never a rant.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "davidnichols-ops/claude-yolo-vibes-v5",
    torch_dtype="auto",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("davidnichols-ops/claude-yolo-vibes-v5")

messages = [
    {"role": "system", "content": "You are claude-yolo-vibes, a coding assistant with a sharp, playful edge."},
    {"role": "user", "content": "Write a Python function that computes fibonacci with memoization."},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Data Provenance & Decontamination

—Primary: Manusagents/GPT-5.5-Gemini-3.1-Pro-Grok-4-Claude-Fable-5-Mythos-5-Qwen-3.7-Max-and-more-Distillation-Dataset (MIT), 16.7M rows; we sampled 59,916 quality-filtered rows (~92% coding by category audit).
—Empty/broken sources (DeepSeek-v4-Pro-Agent, Royal Ghost Coder) were skipped after inspection.
—All sampled rows were checked for 8-gram overlap with HumanEval prompts and excluded on match.
—Personality rows are hand-written (not scraped), so the "voice" is deliberate rather than emergent.

Limitations

—Distillation traces come from other models; behavior on out-of-distribution agentic workloads is inherited from those traces.
—HumanEval is a single-run greedy measurement on a local harness — not EvalPlus-verified.
—The general-domain QA slice is small (5k rows); this is a coding model first.

Also Available / Planned

—Ollama (GGUF Q4_K_M): planned — quantization in progress.
—MLX 4-bit: planned for Apple Silicon.
—v4 family: claude-yolo-vibes-v4-dpo (this model's parent), claude-yolo-vibes-v4-mlx-4bit, claude-yolo-vibes-v4-GGUF.

License

Apache 2.0 (base model: Qwen2.5-Coder-7B-Instruct; training data: MIT).