Modelpublic

huluhuluu/llama-3.2-3b-instruct-eagle3-sharegpt-sw512-epoch9-step231810

sourceHugging Faceotherupdated 9d agoView on Hugging Face

0likes1.1kdownloads

Model Card

Llama-3.2 3b Instruct EAGLE3 ShareGPT SW512

This repository contains a SpecForge EAGLE3 draft-model checkpoint for use with meta-llama/Llama-3.2-3B-Instruct. It is a draft model for speculative decoding, not a standalone target language model.

Checkpoint

Field	Value
Source run	`llama3.2-3b-inst-eagle3-sharegpt-sw512`
Checkpoint	`epoch_9_step_231810`
Epoch	`9`
Global step	`231810`
Files	`config.json`, `model.safetensors`, `training_state.pt` (when present)

Training Parameters

Parameter	Value
Base model	`meta-llama/Llama-3.2-3B-Instruct`
Method	`SpecForge EAGLE3 online training`
Training data	`sharegpt_train_clean.jsonl`
Learning rate	`0.0001`
Batch size	`1`
Target batch size	`1`
Epochs configured	`10`
Max length	`2048`
Warmup ratio	`0.015`
Max grad norm	`0.5`
TTT length	`7`
Draft layers	`Not recorded`
Draft sliding window	`512`
Save interval	`5000`
Eval interval	`5000`
Seed	`0`
TP / DP size	`1 / 4`
Attention backend	`sdpa`
Target model backend	`sglang`
SGLang attention backend	`flashinfer`
Dataset build workers	`64`

Draft Model Configuration

Field	Value
Architecture	`LlamaForCausalLMEagle3`
dtype	`bfloat16`
Hidden size	`3072`
Intermediate size	`8192`
Draft layers	`1`
Attention heads	`24`
KV heads	`8`
Draft vocab size	`32000`
Vocab size	`128256`
Max position embeddings	`131072`
Sliding window	`512`
Max window layers	`Not recorded`
Future hidden	`Not recorded`

Notes

—This is the highest-step local checkpoint available when this repository was published.
—The checkpoint is intended to be loaded by SpecForge/EAGLE3-compatible code.
—training_state.pt is included when available for provenance and training-state inspection.
—No benchmark claim is made in this card.