darwinkernelpanic/luau-codellama-7b-reasoning

See axolotl config

axolotl version: 0.13.0.dev0

base_model: codellama/CodeLlama-7b-hf
model_type: LlamaForCausalLM
tokenizer_type: LlamaTokenizer

# Keep full precision weights (fast on Hopper)
load_in_8bit: false
load_in_4bit: false
strict: false

chat_template: llama3

datasets:
  - path: darwinkernelpanic/luau-reasoning-normalized
    type: chat_template
    conversation: llama3
    field_messages: messages
    add_generation_prompt: true

# Preprocessing workers (CPU). Fine as-is.
num_proc: 16

output_dir: ./outputs/luau-codellama-h200-fast

# ===== LoRA =====
adapter: lora
lora_r: 16
lora_alpha: 32
lora_dropout: 0.05
lora_target_modules:
  - q_proj
  - k_proj
  - v_proj
  - o_proj

# ===== Precision =====
bf16: true
fp16: false
tf32: true

# ===== Sequence / batching =====
sequence_len: 4096
# Keep packing for throughput, but enable length grouping to cut padding
sample_packing: true
group_by_length: true

# Lower micro-batch a bit to kill peak VRAM while staying fast
micro_batch_size: 5
gradient_accumulation_steps: 1

# ===== Training =====
num_epochs: 3
optimizer: adamw_torch
learning_rate: 2e-4
lr_scheduler_type: cosine
warmup_steps: 100

train_on_inputs: false

# Turn on checkpointing — tiny speed hit, big memory win
gradient_checkpointing: true
gradient_clipping: 1.0

# ===== Dataloader =====
# Keep pin_memory, but avoid too many loader workers in Accelerate
dataloader_num_workers: 2
dataloader_pin_memory: true
# Optional: avoid insanely large host->device prefetch
# dataloader_prefetch_factor: 2

# ===== Logging / eval =====
logging_steps: 25
val_set_size: 0.05
# Reduce eval/save frequency to avoid spikes
eval_steps: 1000
save_strategy: steps
save_steps: 1000
save_total_limit: 3

seed: 42

# ===== DeepSpeed =====
# Off for single H200 — overhead not worth it for 7B

outputs/luau-codellama-h200-fast

This model is a fine-tuned version of codellama/CodeLlama-7b-hf on the darwinkernelpanic/luau-reasoning-normalized dataset. It achieves the following results on the evaluation set:

• Loss: 0.4927
• Ppl: 1.6368
• Memory/max Active (gib): 19.1
• Memory/max Allocated (gib): 19.1
• Memory/device Reserved (gib): 139.06

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0002
• train_batch_size: 5
• eval_batch_size: 5
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 100
• training_steps: 3996

Training results

Training Loss	Epoch	Step	Validation Loss	Ppl	Active (gib)	Allocated (gib)	Reserved (gib)
No log	0	0	1.6888	5.4129	18.94	18.94	139.12
0.5511	0.7502	1000	0.5410	1.7177	19.1	19.1	139.02
0.5052	1.5004	2000	0.5064	1.6593	19.1	19.1	139.06
0.4733	2.2506	3000	0.4927	1.6368	19.1	19.1	139.06

Framework versions

• PEFT 0.18.0
• Transformers 4.57.1
• Pytorch 2.8.0+cu128
• Datasets 4.4.1
• Tokenizers 0.22.1