القائمة
Rank candidate prompts by computing LLM perplexity of the full conversation conditioned on each prompt, selecting the lowest-perplexity candidate as the best match
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
Memory-efficient Keras generator that streams sharded CSV files in chunks, renders strokes to images on-the-fly, and yields batches for training on datasets too large for memory
Stack 1D convolutions for local feature extraction before bidirectional LSTMs to classify variable-length stroke sequences into hundreds of doodle categories
Partition a large dataset into N balanced shards using integer key modulo arithmetic for reproducible, class-interleaved splits across CSV files
Normalize raw stroke coordinates to 0-255 range, resample at uniform arc-length spacing, then apply Ramer-Douglas-Peucker simplification
Render stroke sequences to grayscale images with temporal intensity encoding where earlier strokes are brighter and later strokes fade to encode drawing order
Score each face in a video as anomalous by computing L2 distance from the embedding centroid of all faces, then convert to probability via logistic function
| name | llm-perplexity-prompt-ranking |
| description | Rank candidate prompts by computing LLM perplexity of the full conversation conditioned on each prompt, selecting the lowest-perplexity candidate as the best match |
When selecting among multiple candidate prompts (e.g., which instruction produced a given output), format each candidate into a full conversation template with the input/output text, compute the perplexity of each formatted sequence, and rank by lowest perplexity. The LLM assigns lower perplexity to sequences that are more "natural" given its training — the prompt that actually produced the output will typically score lowest.
import torch
import torch.nn as nn
from transformers import AutoModelForCausalLM, AutoTokenizer
def rank_prompts(candidates, context_template, model, tokenizer):
loss_fn = nn.CrossEntropyLoss(reduction='none')
scores = []
texts = [context_template.format(prompt=p) for p in candidates]
inputs = tokenizer(texts, return_tensors="pt", padding=True,
truncation=True, add_special_tokens=False).to(model.device)
with torch.no_grad():
logits = model(**inputs).logits
labels = inputs["input_ids"].clone()
labels[~inputs["attention_mask"].bool()] = -100
shift_logits = logits[:, :-1].contiguous()
shift_labels = labels[:, 1:].contiguous()
for i in range(len(candidates)):
loss = loss_fn(shift_logits[i], shift_labels[i])
valid = (shift_labels[i] != -100).sum()
scores.append((loss.sum() / valid).item())
ranked = sorted(zip(candidates, scores), key=lambda x: x[1])
return ranked[0][0], ranked
best_prompt, all_ranked = rank_prompts(
["Rewrite formally", "Make it rhyme", "Simplify"],
"Instruction: {prompt}\nInput: ...\nOutput: ...",
model, tokenizer)