pandyamarut/Qwen-3-8B-LoRa

See axolotl config

axolotl version: 0.13.0.dev0

# adapter: lora
# base_model: Qwen/Qwen3-8B
# bf16: true
# dataset_prepared_path: last_run_prepared

# # Dataset configuration for instruction/input/output format
# datasets:
# - chat_template: tokenizer_default
#   field_messages: messages
#   message_field_content: content
#   message_field_role: role
#   path: data_clean.jsonl
#   roles:
#     assistant:
#     - assistant
#     system:
#     - system
#     user:
#     - user
#   type: chat_template

# debug: null
# deepspeed: /osmosis/zero2.json
# early_stopping_patience: null
# flash_attention: true
# fp16: null
# fsdp: null
# fsdp_config: null
# gradient_accumulation_steps: 1
# gradient_checkpointing: true
# gradient_checkpointing_kwargs:
#   use_reentrant: false
# group_by_length: false
# learning_rate: 0.0001
# liger_fused_linear_cross_entropy: true
# liger_glu_activation: true
# liger_layer_norm: true
# liger_rms_norm: true
# liger_rope: true
# load_in_4bit: false
# load_in_8bit: false
# logging_steps: 1
# lora_alpha: 64
# lora_dropout: 0.05
# lora_model_dir: null
# lora_r: 64
# lora_target_linear: true
# lr_scheduler: cosine
# max_grad_norm: 1
# micro_batch_size: 16
# model_type: AutoModelForCausalLM
# num_epochs: 1
# optimizer: adamw_torch
# output_dir: ./outQwen
# pad_to_sequence_len: true
# plugins:
# - axolotl.integrations.liger.LigerPlugin
# resume_from_checkpoint: null
# sample_packing: false
# save_steps: 60
# save_total_limit: 100
# sequence_len: 4096
# # special_tokens:
# #   eos_token: <|im_end|>
# strict: false
# tf32: false
# tokenizer_type: AutoTokenizer
# train_on_inputs: false
# trust_remote_code: true
# val_set_size: 0.0
# wandb_entity: test-aa
# wandb_project: seedcoder
# wandb_log_model: null
# wandb_name: qwen-data-pattern
# wandb_watch: null
# warmup_ratio: 0.05
# weight_decay: 0.0
# xformers_attention: null
adapter: lora
base_model: osmosis-ai/sourcegraph-coder-8b-v1
bf16: true
dataset_prepared_path: last_run_prepared

# Dataset configuration for instruction/input/output format
datasets:
- chat_template: tokenizer_default
  field_messages: messages
  message_field_content: content
  message_field_role: role
  path: data_clean.jsonl
  roles:
    assistant:
    - assistant
    system:
    - system
    user:
    - user
  type: chat_template

debug: null
deepspeed: /osmosis/zero2.json
early_stopping_patience: null
flash_attention: true
fp16: null
fsdp: null
fsdp_config: null
gradient_accumulation_steps: 4
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
group_by_length: false
learning_rate: 0.0001
liger_fused_linear_cross_entropy: false
liger_glu_activation: false
liger_layer_norm: false
liger_rms_norm: false
liger_rope: false
load_in_4bit: false
load_in_8bit: false
logging_steps: 1
lora_alpha: 64
lora_dropout: 0.05
lora_model_dir: null
lora_r: 64
lora_target_linear: true
lr_scheduler: cosine
max_grad_norm: 1
micro_batch_size: 4
model_type: AutoModelForCausalLM
num_epochs: 1
optimizer: adamw_torch
output_dir: ./lora-out-qwen38b-sft
pad_to_sequence_len: true
plugins:
- axolotl.integrations.liger.LigerPlugin
resume_from_checkpoint: null
sample_packing: false
save_steps: 60
save_total_limit: 100
sequence_len: 4096
# special_tokens:
#   eos_token: <|im_end|>
strict: false
tf32: false
tokenizer_type: AutoTokenizer
train_on_inputs: false
trust_remote_code: true
val_set_size: 0.0
wandb_entity: test-aa
wandb_project: seedcoder
wandb_log_model: null
wandb_name: qwe-3-8b-sft-run
wandb_watch: null
warmup_ratio: 0.05
weight_decay: 0.0
xformers_attention: null

lora-out-qwen38b-sft

This model is a fine-tuned version of osmosis-ai/sourcegraph-coder-8b-v1 on the data_clean.jsonl dataset.

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: 0.0001
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 4
• total_train_batch_size: 128
• total_eval_batch_size: 32
• 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: 12
• training_steps: 258

Training results

Framework versions

• PEFT 0.17.1
• Transformers 4.57.1
• Pytorch 2.8.0+cu128
• Datasets 4.3.0
• Tokenizers 0.22.1