GKD

GKD (Generalized Knowledge Distillation) training algorithm is proposed in the paper On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes. This algorithm transfers knowledge from the teacher model to the student model by combining offline and on-policy learning strategies.

Loss Function

Given an input sequence \(x\) and output sequence \(y\), the GKD loss function can be written as:

\[ \mathcal{L}_{\text{GKD}}(x, y) = \sum_{t=1}^{|y|} D(P_{\text{teacher}}(\cdot | x, y_{<t}), P_{\text{student}}(\cdot | x, y_{<t})) \]

Where:

\(y_{<t} = (y_1, y_2, \ldots, y_{t-1})\): sequence of the first \(t-1\) tokens
\(P_{\text{teacher}}(\cdot | x, y_{<t})\): output probability distribution of the teacher model given context \(x, y_{<t}\)
\(P_{\text{student}}(\cdot | x, y_{<t})\): output probability distribution of the student model given context \(x, y_{<t}\)
\(D(\cdot, \cdot)\): divergence function to measure the difference between two probability distributions

Divergence Metrics

KL Divergence (Kullback-Leibler Divergence)

KL divergence is an asymmetric measure of the difference between two probability distributions \(P\) and \(Q\):

\[ \text{KL}(P \| Q) = \sum_v P(v) \log \frac{P(v)}{Q(v)} = \mathbb{E}_{v \sim P}\left[\log \frac{P(v)}{Q(v)}\right] \]

Forward KL and Reverse KL

In knowledge distillation, there are two choices depending on the order of the two distributions in the KL divergence:

Forward KL

\[ \text{KL}(P_{\text{teacher}} \| P_{\text{student}}) = \sum_v P_{\text{teacher}}(v) \log \frac{P_{\text{teacher}}(v)}{P_{\text{student}}(v)} \]

Characteristics: Mode-covering

Expectation is computed under the teacher distribution
The student model tends to cover the entire teacher distribution (including low-probability regions)

Reverse KL

\[ \text{KL}(P_{\text{student}} \| P_{\text{teacher}}) = \sum_v P_{\text{student}}(v) \log \frac{P_{\text{student}}(v)}{P_{\text{teacher}}(v)} \]

Characteristics: Mode-seeking

Expectation is computed under the student distribution
The student model tends to concentrate on the peak regions (high-probability areas) of the teacher model

Generalized Jensen-Shannon Divergence (Generalized JSD)

GKD uses generalized JSD as the core metric, performing smooth interpolation between Forward KL and Reverse KL through parameter \(\beta \in [0, 1]\).

For two probability distributions \(P\) and \(Q\), generalized JSD is defined as:

\[ D_{\text{JSD}(\beta)}(P, Q) = \beta \cdot \text{KL}(P \| M) + (1-\beta) \cdot \text{KL}(Q \| M) \]

Where the mixture distribution \(M\) is defined as:

\[ M = \beta \cdot P + (1-\beta) \cdot Q \]

When \(\beta = 0.5\), it reduces to the standard symmetric JSD
By adjusting \(\beta\), one can trade off between Mode-seeking and Mode-covering

In GKD, we set \(P = P_{\text{teacher}}\) and \(Q = P_{\text{student}}\), therefore:

\[ D_{\text{JSD}(\beta)}(P_{\text{teacher}}, P_{\text{student}}) = \beta \cdot \text{KL}(P_{\text{teacher}} \| M) + (1-\beta) \cdot \text{KL}(P_{\text{student}} \| M) \]

Where \(M = \beta \cdot P_{\text{teacher}} + (1-\beta) \cdot P_{\text{student}}\)

For extreme cases (\(\beta = 0\) or \(\beta = 1\)), directly compute a single KL divergence:

When \(\beta = 0\): directly define \(D = \text{KL}(P_{\text{teacher}} \| P_{\text{student}})\) (Forward KL, Mode-covering)

When \(\beta = 1\): directly define \(D = \text{KL}(P_{\text{student}} \| P_{\text{teacher}})\) (Reverse KL, Mode-seeking)

When \(0 < \beta < 1\): use the above mixture distribution formula for interpolation

By adjusting the \(\beta\) parameter, interpolation can be performed between different divergence metrics. When \(\beta = 0.5\), the divergence is the standard symmetric JSD.

Three Training Modes

GKD training has three training modes, distinguished by the source of the output sequence \(y\).

Mode Selection Logic

During training, each sample selects a mode according to the following priority:

# Pseudocode: mode selection logic
if random() < lmbda:
    # Mode 1: On-Policy learning, output sequence sampled by student model
    y = student.generate(x)
    source = "student"
elif seq_kd:
    # Mode 2: Sequential KD, output sequence sampled by teacher model
    y = teacher.generate(x)
    source = "teacher"
else:
    # Mode 3: Offline learning, use output sequence from dataset
    y = y_ground_truth
    source = "dataset"

# Same loss function
loss = D_JSD(P_teacher(·|x,y), P_student(·|x,y))

Mode 1: On-Policy Learning

Set parameter lambda, triggered with probability \(\lambda\), using student model sampling \(y \sim P_{\text{student}}(\cdot | x)\)

The student model learns from sequences generated by itself
Exposed to errors it might make, learning to self-correct and recover from errors
Aligns training distribution with inference distribution
Improves model robustness and practical application performance

Applicable Scenarios:

The student model already has certain generation capabilities
Want to improve model performance in real inference scenarios

Mode 2: Sequential KD (`seq_kd=True` and on-policy not triggered)

Set parameter seq_kd=True, when on-policy is not triggered, use teacher model sampling

Data Source: \(y \sim P_{\text{teacher}}(\cdot | x)\)

Mode 3: Offline Learning (other cases)

Data Source: \(y = y^* \sim \text{Dataset}\)

The student model learns from annotated sequences in the dataset

Parameter Settings

We can perform GKD training by setting the following parameters:

Parameter	Type	Default	Range	Description
`--teacher_model`	str	Required	-	Teacher model path or model ID
`--beta`	float	0.5	[0.0, 1.0]	Divergence interpolation coefficient • 0.0: Forward KL • 0.5: JSD (balanced) • 1.0: Reverse KL
`--lmbda`	float	0.5	[0.0, 1.0]	On-Policy learning trigger probability • 0.0: Pure Offline • 0.5: Mixed strategy (recommended) • 1.0: Pure On-Policy
`--seq_kd`	bool	False	True/False	Whether to use teacher-generated sequences • False: Use dataset when not on-policy • True: Use teacher generation when not on-policy
`--temperature`	float	0.9	> 0	Generation sampling temperature, controls randomness
`--sft_alpha`	float	0	>= 0	Mix in a proportion of SFT loss; applied to non-student-generated completions
`--max_completion_length`	int	512	> 0	Maximum number of tokens during generation

Sampling Acceleration

In GKD training, there are two types of online sampling scenarios:

Student model sampling (when lmbda > 0): triggered with probability \(\lambda\)
Teacher model sampling (when seq_kd=True): triggered with probability \(1-\lambda\)

Since the sampling process significantly slows down training speed, you can refer to the following two acceleration schemes:

Solution 1: Student Model Sampling Acceleration

Requirement: swift >= 3.10.dev

Use vLLM as the inference backend to accelerate student model sampling. Supports two deployment modes, consistent with GRPO. Refer to GRPO documentation

Note: vLLM acceleration only applies to student model on-policy sampling (lmbda > 0). Teacher model sequential KD sampling (seq_kd=True) currently still uses PyTorch. Pre-sampling scheme is recommended.

Training script reference here, for related parameters, please refer to GRPO vLLM Parameters.

Solution 2: Teacher Model Pre-sampling

For teacher model sampling (seq_kd=True), pre-sampling is recommended: first use the teacher model to offline generate high-quality data, then train.

Step 1: Generate data using teacher model

export teacher_model='OpenGVLab/InternVL3-8B'

NPROC_PER_NODE=4 \
CUDA_VISIBLE_DEVICES=0,1,2,3 \
swift infer \
    --model $teacher_model \
    --infer_backend vllm \
    --val_dataset 'modelscope/coco_2014_caption:validation#5000' \
    --vllm_gpu_memory_utilization 0.9 \
    --vllm_max_model_len 8192 \
    --max_new_tokens 2048 \
    --write_batch_size 1000 \
    --result_path teacher_generated_data.jsonl

Step 2: Train using pre-generated data

swift rlhf \
    --rlhf_type gkd \
    --model OpenGVLab/InternVL3-2B-Pretrained \
    --teacher_model $teacher_model \
    --dataset 'teacher_generated_data.jsonl' \
    --seq_kd false \
    ...

Training script reference here

On-Policy Distillation

We can achieve the On-Policy Distillation training described in the Thinking Machines Lab blog by setting the following parameters:

--lmbda 1 # on-policy
--beta 1 # reverse

For a complete implementation, refer to the example script here.