DeepSeek-V3-DRAFT-0.6B-v3.0

A 0.6B parameter draft (speculative decoding) model for use with DeepSeek-V3-0324 and DeepSeek-V3.

See DeepSeek-V3-DRAFT-0.6B-v3.0-GGUF for the models in gguf format for use with llama.cpp.

---

Extending the context above 32k

The current config.json is set for context length up to 32k tokens. Add the "rope_scaling" section to config.json to enable YaRN, eg:

To extend the context to 64k:

  "max_position_embeddings": 65536,
  ...
  "rope_scaling": {
    "factor": 2.0,
    "original_max_position_embeddings": 32768,
    "type": "yarn"
  },

To extend the context to 128k:

  "max_position_embeddings": 131072,
  ...
  "rope_scaling": {
    "factor": 4.0,
    "original_max_position_embeddings": 32768,
    "type": "yarn"
  },

To extend the context to 160k:

  "max_position_embeddings": 163840,
  ...
  "rope_scaling": {
    "factor": 5.0,
    "original_max_position_embeddings": 32768,
    "type": "yarn"
  },

NOTE: Because llama.cpp uses "static-YaRN" the scaling factor remains constant regardless of input length! Only add the rope_scaling configuration when processing long contexts is required...

---

How this model was created

1. The initial model was created from Qwen2.5-0.5B-Instruct using transplant-vocab:

> python ./transplant_vocab.py \
    ./Qwen2.5-0.5B-Instruct \
    ./DeepSeek-V3-0324 \
    ./DeepSeek-V3-DRAFT-0.6B-UNTRAINED \
    --override "<｜begin▁of▁sentence｜>" "<|endoftext|>" \
    --override "<｜end▁of▁sentence｜>" "<|im_end|>" \
    --override "<｜▁pad▁｜>" "<|endoftext|>" \
	--override "<｜fim▁hole｜>" "<|fim_middle|>" \
	--override "<｜fim▁begin｜>" "<|fim_prefix|>" \
	--override "<｜fim▁end｜>" "<|fim_suffix|>" \
	--override "<｜User｜>" "<|im_start|>user\\n" \
	--override "<｜Assistant｜>" "<|im_start|>assistant\\n" \
	--override "<|EOT|>" "<|endoftext|>" \
	--override "<｜tool▁calls▁begin｜>" "<tool_call>" \
	--override "<｜tool▁calls▁end｜>" "</tool_call>" \
	--override "<｜tool▁call▁begin｜>" "<tool_call>" \
	--override "<｜tool▁call▁end｜>" "</tool_call>" \
	--override "<｜tool▁outputs▁begin｜>" "<tool_response>" \
	--override "<｜tool▁outputs▁end｜>" "</tool_response>" \
	--override "<｜tool▁output▁begin｜>" "<tool_response>" \
	--override "<｜tool▁output▁end｜>" "</tool_response>" \
	--override "<｜tool▁sep｜>" "</tool_call>"

Loading config from 'Qwen2.5-0.5B-Instruct'... Done.
Loading config from 'DeepSeek-V3-0324'... Done.
Loading tokenizer from 'Qwen2.5-0.5B-Instruct'... Done.
Loading tokenizer from 'DeepSeek-V3-0324'... Done.
Loading model from 'Qwen2.5-0.5B-Instruct'... Done.

Input model configuration:
- Target vocabulary size    : 129280 (used = 128815, unused = 465)
- Donor vocabulary size     : 151936
- Donor num layers          : 24 (tied embeddings = True)
- Donor hidden size         : 896
- Donor attention heads     : 14
- Donor intermediate size   : 4864 (ratio = 1:5.4)
- Donor total parameters    : 494032768 (0.49B)
-- Embedding parameters     : 136134656 (0.14B)
-- Non-embedding parameters : 357898112 (0.36B)

Processing 3 automatic token overrides:
✔ 'bos_token_id' : 0 '<｜begin▁of▁sentence｜>' → [151643] '<|endoftext|>'
✔ 'eos_token_id' : 1 '<｜end▁of▁sentence｜>' → [151645] '<|im_end|>'
✘ 'pad_token_id' : 1 is already mapped to [151645]

Processing 18 manual token overrides:
✔      0 : '<｜begin▁of▁sentence｜>' → [151643] '<|endoftext|>'
✔      1 : '<｜end▁of▁sentence｜>' → [151645] '<|im_end|>'
✔      2 : '<｜▁pad▁｜>' → [151643] '<|endoftext|>'
✔ 128800 : '<｜fim▁hole｜>' → [151660] '<|fim_middle|>'
✔ 128801 : '<｜fim▁begin｜>' → [151659] '<|fim_prefix|>'
✔ 128802 : '<｜fim▁end｜>' → [151661] '<|fim_suffix|>'
✔ 128803 : '<｜User｜>' → [151644, 872, 198] '<|im_start|>user\n'
✔ 128804 : '<｜Assistant｜>' → [151644, 77091, 198] '<|im_start|>assistant\n'
✔ 128805 : '<|EOT|>' → [151643] '<|endoftext|>'
✔ 128806 : '<｜tool▁calls▁begin｜>' → [151657] '<tool_call>'
✔ 128807 : '<｜tool▁calls▁end｜>' → [151658] '</tool_call>'
✔ 128808 : '<｜tool▁call▁begin｜>' → [151657] '<tool_call>'
✔ 128809 : '<｜tool▁call▁end｜>' → [151658] '</tool_call>'
✔ 128810 : '<｜tool▁outputs▁begin｜>' → [27, 14172, 9655, 29] '<tool_response>'
✔ 128811 : '<｜tool▁outputs▁end｜>' → [522, 14172, 9655, 29] '</tool_response>'
✔ 128812 : '<｜tool▁output▁begin｜>' → [27, 14172, 9655, 29] '<tool_response>'
✔ 128813 : '<｜tool▁output▁end｜>' → [522, 14172, 9655, 29] '</tool_response>'
✔ 128814 : '<｜tool▁sep｜>' → [151658] '</tool_call>'

NOTE: Using an "untied" copy of 'embed_tokens.weight' as new 'lm_head.weight' tensor...

Transplanting tokens: 100%|████████████████████████████████████████████████████████████| 128815/128815 [00:53<00:00, 2423.79token/s]

Transplant mappings:
- 1 to 1  : 83683 (65%)
- 2 to 1  : 38380 (30%)
- 3 to 1  : 4583 (3.6%)
- 4 to 1  : 927 (0.72%)
- 5 to 1  : 273 (0.21%)
- 6 to 1  : 91 (0.071%)
- 7 to 1  : 35 (0.027%)
- 8 to 1  : 22 (0.017%)
- 9 to 1  : 8 (0.0062%)
- 10 to 1 : 4 (0.0031%)
- 11 to 1 : 4 (0.0031%)
- 13 to 1 : 1 (0.00078%)
- 14 to 1 : 10 (0.0078%)
- 15 to 1 : 91 (0.071%)
- 16 to 1 : 701 (0.54%)
- 19 to 1 : 1 (0.00078%)
- 21 to 1 : 1 (0.00078%)

Head initialized with:
- Copies : 83683 (65%)
- Means  : 45132 (35%)
- Zeros  : 465 (0.36%)

Output model configuration:
- Output vocabulary size    : 129280
- Output num layers         : 24 (tied embeddings = False)
- Output hidden size        : 896
- Output attention heads    : 14
- Output intermediate size  : 4864 (ratio = 1:5.4)
- Output total parameters   : 589567872 (0.59B)
-- Embedding parameters     : 231669760 (0.23B)
-- Non-embedding parameters : 357898112 (0.36B)

Saving model and tokenizer to 'DeepSeek-V3-DRAFT-0.6B-UNTRAINED' folder
[2025-08-07 15:36:33,693] [INFO] [real_accelerator.py:203:get_accelerator] Setting ds_accelerator to cuda (auto detect)

Patching 'torch_dtype' in 'DeepSeek-V3-DRAFT-0.6B-UNTRAINED/config.json' based on actual saved tensors
- Updated 'torch_dtype' to 'bfloat16' based on actual tensor dtype

Operation completed successfully (ignore any 'segmentation fault' that follows!!!)

2. The following datasets were used to create a fine-tuning dataset of ~2.3B tokens:

• agentlans/common-crawl-sample
• bigcode/the-stack-smol-xl
• rombodawg/Everything_Instruct (NOTE: output field only)

formatted just between <｜end▁of▁sentence｜> tags.

3. The model was then trained using qlora-pipe-lite for 1 epoch with a batch size of 60 and a sequence length of 32k (~2M tokens per step):

# ==============================
# MODEL AND OUTPUT CONFIGURATION
# ==============================

model_dir = 'models/DeepSeek-V3-DRAFT-0.6B-UNTRAINED'
output_dir = 'finetuned'

# ===========================
# TRAINING TYPE CONFIGURATION
# ===========================

full_fine_tune = true

# =======================
# OPTIMIZER CONFIGURATION
# =======================

lr = 5e-5

# ======================
# TRAINING CONFIGURATION
# ======================

sequence_len = 32768

gradient_accumulation_steps = 10  # 10×6 = batch size 60, 10×6×32768 = ~2M tokens per step

# =====================
# DATASET CONFIGURATION
# =====================

[[datasets]]
dataset_path = 'datasets/common-crawl-sample/*.json'
drop_tails = true

[[datasets]]
dataset_path = 'datasets/the-stack-smol-xl/*.jsonl'
drop_tails = true

[[datasets]]
dataset_path = 'datasets/rombodawg-Everything-Instruct/*.json'
drop_tails = true

I used six RTX A6000 GPUs over three nodes and hence the 60 batch size (6 x 10 gradient accumulation steps = 60):