▸case-18 During a 10-epoch instruction fine-tuning run on Hugging Face `Trainer`, the training loss steadily drops from 2.5 to 0.1, but after epoch 3, evaluation loss starts climbing from 1.1 up to 1.8. What issue is occurring and what Hugging Face trainer callback mechanism should be applied? | pass→pass | 6,168 | 6,238 | +1% | 1 | 1 | 0% | 1,342 | 1,507 | +12% | 0 | 0 | — |
▸case-01 Our engineering team needs to extract standardized JSON entities (invoice number, date, vendor, total) from incoming vendor emails. We have a static set of 5 prompt guidelines and 100 sample emails. The team is considering fine-tuning Llama-3-8B on GPU instances to learn this format. What architecture or approach should we adopt for this static instruction-following task? | pass→fail | 15,565 | 18,033 | +16% | 1 | 1 | 0% | 2,725 | 3,616 | +33% | 0 | 0 | — |
▸case-02 Our company maintains an internal knowledge base of 15,000 Confluence pages that updates multiple times every day. The product team proposed fine-tuning our open-source LLM every night on the updated documents so the model knows current company factual information. What system design should we implement? | pass→pass | 15,303 | 14,729 | -4% | 1 | 1 | 0% | 2,494 | 2,815 | +13% | 0 | 0 | — |
▸case-03 We have an already fine-tuned Llama-3-8B model artifact stored in Hugging Face SafeTensors format. We need to deploy this existing model to serve 200 concurrent HTTP request streams with high token generation throughput and continuous batching on a single NVIDIA A10G GPU. Provide the deployment architecture and serving framework configuration for hosting this model. | pass→pass | 18,195 | 20,652 | +14% | 1 | 1 | 0% | 3,695 | 4,246 | +15% | 0 | 0 | — |
▸case-04 We are setting up a 4-bit QLoRA training script for Llama-3-8B on an NVIDIA A10G (24GB VRAM) using Hugging Face transformers and peft. We want to maximize memory efficiency while preserving normal float quantization quality over standard FP4. Write the python code configuring `BitsAndBytesConfig` for loading the base model. | pass→pass | 5,398 | 6,200 | +15% | 1 | 1 | 0% | 1,202 | 1,488 | +24% | 0 | 0 | — |
▸case-05 When fine-tuning Llama-3-8B using LoRA in Hugging Face PEFT, developers often default to setting `target_modules=["q_proj", "v_proj"]` to save adapter parameters. Provide the recommended `LoraConfig` target module selection to achieve optimal model performance and capacity across projection layers. | pass→pass | 7,802 | 9,032 | +16% | 1 | 1 | 0% | 1,601 | 2,091 | +31% | 0 | 0 | — |
▸case-06 In Hugging Face PEFT `LoraConfig`, we set rank `r=16` for fine-tuning a 7B parameter model. Beginners often set `lora_alpha=1` or `lora_alpha=256`. Provide the standard formula and recommended numerical setting for `lora_alpha` relative to rank `r`. | pass→pass | 8,046 | 6,026 | -25% | 1 | 1 | 0% | 1,683 | 1,534 | -9% | 0 | 0 | — |
▸case-07 We are preparing a dataset for Direct Preference Optimization using Hugging Face TRL `DPOTrainer`. A developer formatted the dataset JSONL with fields `{"instruction": "...", "response": "...", "rating": 5}`. Provide the correct schema keys required by `DPOTrainer` for preference alignment. | pass→pass | 8,576 | 6,201 | -28% | 1 | 1 | 0% | 1,740 | 1,475 | -15% | 0 | 0 | — |
▸case-08 We are building a multi-turn conversational dataset in JSONL format for instruction fine-tuning Llama models using Hugging Face SFTTrainer. A teammate proposed putting raw prompt/response text pairs in string fields `{"input": "...", "output": "..."}`. Show the standardized message object schema needed for multi-turn chat templates. | pass→pass | 9,189 | 10,715 | +17% | 1 | 1 | 0% | 2,003 | 2,252 | +12% | 0 | 0 | — |
▸case-09 We are switching from full parameter fine-tuning (where we used a learning rate of 2e-5) to parameter-efficient fine-tuning with QLoRA on an 8B parameter model. Should we keep the learning rate at 2e-5 or adjust it? Provide the recommended learning rate range for QLoRA/LoRA fine-tuning. | pass→pass | 8,779 | 9,937 | +13% | 1 | 1 | 0% | 1,828 | 2,192 | +20% | 0 | 0 | — |
▸case-10 We are fine-tuning Llama-3-8B on 2 GPUs with 24GB VRAM each. Due to VRAM constraints, `per_device_train_batch_size` must be set to 2. However, target effective global batch size is 64 sequences. Provide the exact setting for `gradient_accumulation_steps` in Hugging Face `TrainingArguments`. | pass→pass | 2,873 | 3,148 | +10% | 1 | 1 | 0% | 678 | 870 | +28% | 0 | 0 | — |
▸case-11 After fine-tuning Llama-3-8B, we want to export the model to GGUF format for local deployment on llama.cpp. A teammate suggested standard Q4_0 uniform 4-bit quantization to minimize file size. Provide the recommended GGUF quantization scheme that preserves model quality while maintaining 4-bit or 5-bit efficiency. | pass→pass | 12,536 | 11,578 | -8% | 1 | 1 | 0% | 2,576 | 2,588 | +0% | 0 | 0 | — |
▸case-12 We are running full parameter fine-tuning of a 70B parameter LLM across 4 NVIDIA A100 (80GB VRAM) GPUs using DeepSpeed. Standard ZeRO Stage 2 runs out of CUDA memory during the forward/backward pass because model parameters and optimizer states exceed total VRAM. Provide the DeepSpeed configuration level and offloading settings required. | pass→pass | 11,102 | 13,109 | +18% | 1 | 1 | 0% | 2,410 | 3,095 | +28% | 0 | 0 | — |
▸case-13 During instruction fine-tuning on a small 2,000 sample chat dataset, our model quickly overfits and generates repetitive, rigid multi-turn responses. Rather than adding severe weight decay, what embedding noise technique can be added to `SFTTrainer` to improve model generalization and conversational fluency, and what parameter controls it? | pass→pass | 5,950 | 6,396 | +7% | 1 | 1 | 0% | 1,195 | 1,465 | +23% | 0 | 0 | — |
▸case-14 When setting up `LoraConfig` for fine-tuning a model on a very small dataset (500 custom domain examples), setting `lora_dropout=0.0` causes rapid overfitting on training examples. What `lora_dropout` value should be configured in `LoraConfig` for small datasets to introduce regularization? | pass→pass | 7,695 | 5,531 | -28% | 1 | 1 | 0% | 1,654 | 1,288 | -22% | 0 | 0 | — |
▸case-15 When configuring Hugging Face `TrainingArguments` for instruction fine-tuning, choosing a constant learning rate schedule often leads to training instability early on and sub-optimal convergence at the end. Specify the recommended learning rate scheduler type and warmup configuration. | pass→pass | 6,248 | 6,148 | -2% | 1 | 1 | 0% | 1,265 | 1,277 | +1% | 0 | 0 | — |
▸case-16 In instruction fine-tuning with Hugging Face TRL `SFTTrainer`, many sample sequences in our dataset are short (200 tokens), but `max_seq_length` is 2048. Standard batching pads every sequence with zeroes, resulting in 80% wasted GPU memory and slow throughput. What `SFTTrainer` parameter concatenates multiple short samples into a single full context window? | pass→pass | 4,541 | 3,973 | -13% | 1 | 1 | 0% | 970 | 916 | -6% | 0 | 0 | — |
▸case-17 We are configuring Hugging Face TRL `DPOTrainer` for preference alignment. An engineer set the DPO loss temperature parameter `beta=1.0`, but the policy model's output distribution drifted drastically from the reference model, degrading generation quality. What is the recommended range for the `beta` parameter in DPO? | fail→pass | 8,046 | 10,698 | +33% | 1 | 1 | 0% | 1,476 | 2,275 | +54% | 0 | 0 | — |
▸case-19 We need to compress a 13B parameter fine-tuned LLM to 4-bit weights for deployment using post-training quantization. Standard Round-To-Nearest (RTN) quantization significantly drops perplexity because it degrades salient activation channels. What activation-aware 4-bit post-training quantization method should be used? | pass→pass | 11,405 | 11,658 | +2% | 1 | 1 | 0% | 2,188 | 2,357 | +8% | 0 | 0 | — |
▸case-20 When loading a Llama-3 model via `AutoModelForCausalLM.from_pretrained` in Hugging Face Transformers for fine-tuning on sequence lengths of 4096 tokens, GPU VRAM usage and attention compute time grow quadratically. What parameter should be passed during model loading to utilize memory-efficient exact attention kernel acceleration? | pass→pass | 4,774 | 5,691 | +19% | 1 | 1 | 0% | 1,042 | 1,429 | +37% | 0 | 0 | — |
▸case-21 We are fine-tuning a Llama model on a single NVIDIA T4 GPU with strict memory limits (16GB VRAM) and slow execution times. Standard Hugging Face PEFT + PyTorch runs out of memory or takes 12 hours. What specialized open-source library and wrapper class provides optimized handwritten CUDA kernels for faster QLoRA training with reduced VRAM footprint? | pass→pass | 6,918 | 8,108 | +17% | 1 | 1 | 0% | 1,633 | 1,858 | +14% | 0 | 0 | — |
▸case-22 During full parameter or adapter fine-tuning of a 14B model, PyTorch throws `torch.cuda.OutOfMemoryError` during the backward pass when computing gradients for intermediate layer activations, even though batch size per GPU is 1. What Hugging Face `TrainingArguments` setting trades a small amount of compute re-evaluation time to drastically reduce activation VRAM memory usage? | pass→pass | 3,517 | 4,976 | +41% | 1 | 1 | 0% | 653 | 1,012 | +55% | 0 | 0 | — |