qwen3-0.6B-reasoning-sft

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>

<details><summary>See axolotl config</summary>

axolotl version: 0.10.0.dev0

# Model Configuration
base_model: Qwen/Qwen3-0.6B-Base
type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
special_tokens:
flash_attention: true
flash_attn_rms_norm: true
flash_attn_fuse_qkv: false
sequence_len: 4096
chat_template: qwen3

# Dataset Configuration
shuffle_merged_datasets: true
dataset_processes: 8
sample_packing: true
pad_to_sequence_len: true
group_by_length: false
train_on_inputs: false
datasets:
  - path: timarni/s1k_r1_clean
    ds_type: json
    type: chat_template
    field_messages: conversations
    message_property_mappings: {role: from, content: value}

# datasets:
#   - path: "json"
#     data_files: "/mloscratch/users/arni/reasoning_sft/data/s1k_r1/s1k_r1_think_token_cleaned.jsonl"
#     type: chat_template
#     ds_type: json
#     split: train
#     field_messages: conversations
#     message_field_role: from
#     message_field_content: value

# Training Hyperparameters
micro_batch_size: 1
gradient_accumulation_steps: 2
max_steps: 8500
num_epochs: 5
learning_rate: 1e-6 # 7e-7
optimizer: adamw_torch
optim_args:
  fused: true
lr_scheduler: cosine
cosine_min_lr_ratio: 0
warmup_ratio: 0.05
weight_decay: 1.0e-4
adam_beta1: 0.9
adam_beta2: 0.95
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
max_grad_norm: 1.0

# Hardware/Performance Configuration
load_in_4bit: false
load_in_8bit: false
deepspeed: /mloscratch/users/arni/meditron_protocol/training/sft/axolotl_config/deepspeed.json
xformers_attention: null
eager_attention: true
tf32: false
bf16: true

# Logging/Checkpointing
output_dir: /mloscratch/users/arni/models/qwen3-0.6B-Base-s1k_r1_reasoning_token
logging_steps: 1
saves_per_epoch: 1
resume_from_checkpoint: null
load_best_model_at_end: false
early_stopping_patience: 0
eval_set_size: 0.0
eval_table_size: null
# evals_per_epoch: 2
# eval_steps: 1000
# save_steps: 100

# WandB Configuration
wandb_project: mnlp # meditron-reasoning
wandb_entity: tim-arni # alexs-team
wandb_name: qwen3-0.6B-Base-s1k_r1_reasoning_token # medicouenne-7b-checkpoint-5742-medMCQA

</details><br>

mloscratch/users/arni/models/qwen3-0.6B-Base-s1k_r1_reasoning_token

This model is a fine-tuned version of Qwen/Qwen3-0.6B-Base on the /mloscratch/users/arni/reasoning_sft/data/s1k_r1/s1k_r1_think_token_cleaned.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: 1e-06
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 2
• total_train_batch_size: 2
• optimizer: Use adamw_torch with betas=(0.9,0.95) and epsilon=1e-08 and optimizer_args=fused=True
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 11
• training_steps: 8500

Training results

Framework versions

• Transformers 4.51.3
• Pytorch 2.6.0+cu124
• Datasets 3.5.0
• Tokenizers 0.21.1