fine-tuning-expert

Name: fine-tuning-expert
Author: Jeffallan

3.0kinstalls971forks

How to install fine-tuning-expert?

jeffallan/claude-skills/fine-tuning-expert

$npx -y skills add Jeffallan/claude-skills --skill fine-tuning-expert

Installs into the current project.

Summary

Use when fine-tuning LLMs, training custom models, or adapting foundation models for specific tasks. Invoke for configuring LoRA/QLoRA adapters, preparing JSONL training datasets, setting hyperparameters for fine-tuning runs, adapter training, transfer learning, finetuning with H

Files · 1

View on GitHub

SKILL.md

1	# Fine-Tuning Expert
2
3	Senior ML engineer specializing in LLM fine-tuning, parameter-efficient methods, and production model optimization.
4
5	## Core Workflow
6
7	1. Dataset preparation — Validate and format data; run quality checks before training starts
8	- Checkpoint: `python validate_dataset.py --input data.jsonl` — fix all errors before proceeding
9	2. Method selection — Choose PEFT technique based on GPU memory and task requirements
10	- Use LoRA for most tasks; QLoRA (4-bit) when GPU memory is constrained; full fine-tune only for small models
11	3. Training — Configure hyperparameters, monitor loss curves, checkpoint regularly
12	- Checkpoint: validation loss must decrease; plateau or increase signals overfitting
13	4. Evaluation — Benchmark against the base model; test on held-out set and edge cases
14	- Checkpoint: collect perplexity, task-specific metrics (BLEU/ROUGE), and latency numbers
15	5. Deployment — Merge adapter weights, quantize, measure inference throughput before serving
16
17	## Reference Guide
18
19	Load detailed guidance based on context:
20
21	\| Topic \| Reference \| Load When \|
22	\|-------\|-----------\|-----------\|
23	\| LoRA/PEFT \| `references/lora-peft.md` \| Parameter-efficient fine-tuning, adapters \|
24	\| Dataset Prep \| `references/dataset-preparation.md` \| Training data formatting, quality checks \|
25	\| Hyperparameters \| `references/hyperparameter-tuning.md` \| Learning rates, batch sizes, schedulers \|
26	\| Evaluation \| `references/evaluation-metrics.md` \| Benchmarking, metrics, model comparison \|
27	\| Deployment \| `references/deployment-optimization.md` \| Model merging, quantization, serving \|
28
29	## Minimal Working Example — LoRA Fine-Tuning with Hugging Face PEFT
30
31	```python
32	from datasets import load_dataset
33	from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments
34	from peft import LoraConfig, get_peft_model, TaskType
35	from trl import SFTTrainer
36	import torch
37
38	# 1. Load base model and tokenizer
39	model_id = "meta-llama/Llama-3-8B"
40	tokenizer = AutoTokenizer.from_pretrained(model_id)
41	tokenizer.pad_token = tokenizer.eos_token
42
43	model = AutoModelForCausalLM.from_pretrained(
44	model_id,
45	torch_dtype=torch.bfloat16,
46	device_map="auto",
47	)
48
49	# 2. Configure LoRA adapter
50	lora_config = LoraConfig(
51	task_type=TaskType.CAUSAL_LM,
52	r=16, # rank — increase for more capacity, decrease to save memory
53	lora_alpha=32, # scaling factor; typically 2× rank
54	target_modules=["q_proj", "v_proj"],
55	lora_dropout=0.05,
56	bias="none",
57	)
58	model = get_peft_model(model, lora_config)
59	model.print_trainable_parameters() # verify: should be ~0.1–1% of total params
60
61	# 3. Load and format dataset (Alpaca-style JSONL)
62	dataset = load_dataset("json", data_files={"train": "train.jsonl", "test": "test.jsonl"})
63
64	def format_prompt(example):
65	return {"text": f"### Instruction:\n{example['instruction']}\n\n### Response:\n{example['output']}"}
66
67	dataset = dataset.map(format_prompt)
68
69	# 4. Training arguments
70	training_args = TrainingArguments(
71	output_dir="./checkpoints",
72	num_train_epochs=3,
73	per_device_train_batch_size=4,
74	gradient_accumulation_steps=4, # effective batch size = 16
75	learning_rate=2e-4,
76	lr_scheduler_type="cosine",
77	warmup_ratio=0.03, # always use warmup
78	fp16=False,
79	bf16=True,
80	logging_steps=10,
81	eval_strategy="steps",
82	eval_steps=100,
83	save_steps=200,
84	load_best_model_at_end=True,
85	)
86
87	# 5. Train
88	trainer = SFTTrainer(
89	model=model,
90	args=training_args,
91	train_dataset=dataset["train"],
92	eval_dataset=dataset["test"],
93	dataset_text_field="text",
94	max_seq_length=2048,
95	)
96	trainer.train()
97
98	# 6. Save adapter weights only
99	model.save_pretrained("./lora-adapter")
100	tokenizer.save_pretrained("./lora-adapter")
101	```
102
103	QLoRA variant — add these lines before loading the model to enable 4-bit quantization:
104	```python
105	from transformers import BitsAndBytesConfig
106
107	bnb_config = BitsAndBytesConfig(
108	load_in_4bit=True,
109	bnb_4bit_quant_type="nf4",
110	bnb_4bit_compute_d

People who liked this also viewed

6 picks

Type

fine-tuning-expert

3.0kinstalls971forks

View on GitHub

How to install fine-tuning-expert?

jeffallan/claude-skills/fine-tuning-expert

$npx -y skills add Jeffallan/claude-skills --skill fine-tuning-expert

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Fine-Tuning Expert
2
3	Senior ML engineer specializing in LLM fine-tuning, parameter-efficient methods, and production model optimization.
4
5	## Core Workflow
6
7	1. Dataset preparation — Validate and format data; run quality checks before training starts
8	- Checkpoint: `python validate_dataset.py --input data.jsonl` — fix all errors before proceeding
9	2. Method selection — Choose PEFT technique based on GPU memory and task requirements
10	- Use LoRA for most tasks; QLoRA (4-bit) when GPU memory is constrained; full fine-tune only for small models
11	3. Training — Configure hyperparameters, monitor loss curves, checkpoint regularly
12	- Checkpoint: validation loss must decrease; plateau or increase signals overfitting
13	4. Evaluation — Benchmark against the base model; test on held-out set and edge cases
14	- Checkpoint: collect perplexity, task-specific metrics (BLEU/ROUGE), and latency numbers
15	5. Deployment — Merge adapter weights, quantize, measure inference throughput before serving
16
17	## Reference Guide
18
19	Load detailed guidance based on context:
20
21	\| Topic \| Reference \| Load When \|
22	\|-------\|-----------\|-----------\|
23	\| LoRA/PEFT \| `references/lora-peft.md` \| Parameter-efficient fine-tuning, adapters \|
24	\| Dataset Prep \| `references/dataset-preparation.md` \| Training data formatting, quality checks \|
25	\| Hyperparameters \| `references/hyperparameter-tuning.md` \| Learning rates, batch sizes, schedulers \|
26	\| Evaluation \| `references/evaluation-metrics.md` \| Benchmarking, metrics, model comparison \|
27	\| Deployment \| `references/deployment-optimization.md` \| Model merging, quantization, serving \|
28
29	## Minimal Working Example — LoRA Fine-Tuning with Hugging Face PEFT
30
31	```python
32	from datasets import load_dataset
33	from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments
34	from peft import LoraConfig, get_peft_model, TaskType
35	from trl import SFTTrainer
36	import torch
37
38	# 1. Load base model and tokenizer
39	model_id = "meta-llama/Llama-3-8B"
40	tokenizer = AutoTokenizer.from_pretrained(model_id)
41	tokenizer.pad_token = tokenizer.eos_token
42
43	model = AutoModelForCausalLM.from_pretrained(
44	model_id,
45	torch_dtype=torch.bfloat16,
46	device_map="auto",
47	)
48
49	# 2. Configure LoRA adapter
50	lora_config = LoraConfig(
51	task_type=TaskType.CAUSAL_LM,
52	r=16, # rank — increase for more capacity, decrease to save memory
53	lora_alpha=32, # scaling factor; typically 2× rank
54	target_modules=["q_proj", "v_proj"],
55	lora_dropout=0.05,
56	bias="none",
57	)
58	model = get_peft_model(model, lora_config)
59	model.print_trainable_parameters() # verify: should be ~0.1–1% of total params
60
61	# 3. Load and format dataset (Alpaca-style JSONL)
62	dataset = load_dataset("json", data_files={"train": "train.jsonl", "test": "test.jsonl"})
63
64	def format_prompt(example):
65	return {"text": f"### Instruction:\n{example['instruction']}\n\n### Response:\n{example['output']}"}
66
67	dataset = dataset.map(format_prompt)
68
69	# 4. Training arguments
70	training_args = TrainingArguments(
71	output_dir="./checkpoints",
72	num_train_epochs=3,
73	per_device_train_batch_size=4,
74	gradient_accumulation_steps=4, # effective batch size = 16
75	learning_rate=2e-4,
76	lr_scheduler_type="cosine",
77	warmup_ratio=0.03, # always use warmup
78	fp16=False,
79	bf16=True,
80	logging_steps=10,
81	eval_strategy="steps",
82	eval_steps=100,
83	save_steps=200,
84	load_best_model_at_end=True,
85	)
86
87	# 5. Train
88	trainer = SFTTrainer(
89	model=model,
90	args=training_args,
91	train_dataset=dataset["train"],
92	eval_dataset=dataset["test"],
93	dataset_text_field="text",
94	max_seq_length=2048,
95	)
96	trainer.train()
97
98	# 6. Save adapter weights only
99	model.save_pretrained("./lora-adapter")
100	tokenizer.save_pretrained("./lora-adapter")
101	```
102
103	QLoRA variant — add these lines before loading the model to enable 4-bit quantization:
104	```python
105	from transformers import BitsAndBytesConfig
106
107	bnb_config = BitsAndBytesConfig(
108	load_in_4bit=True,
109	bnb_4bit_quant_type="nf4",
110	bnb_4bit_compute_d

People who liked this also viewed

6 picks

Type