Sera-4.6-Lite-T2-v4-316-axolotl__Qwen3-8B-v2/README.md

library_name, base_model, tags, datasets, model-index

library_name	base_model	tags	datasets	model-index
transformers	Qwen/Qwen3-8B	generated_from_trainer	laion/Sera-4.6-Lite-T2-v4-316	name results e/data1/datasets/playground/ot-baf/checkpoints/sera-v4-316-axolotl__Qwen3-8B

See axolotl config

axolotl version: 0.16.0.dev0

# Sera v4 axolotl config template — consumes laion/Sera-4.6-Lite-T2-v4-<SIZE>
# where tool_calls are already pre-rendered into content as <tool_call>...</tool_call>
# (Hermes/Qwen3 wire format) per SERA's transform_traj_hermes.
#
# Fill 316 via sed-substitution.

base_model: Qwen/Qwen3-8B
deepspeed: /e/scratch/jureap59/feuer1/code/axolotl/deepspeed_configs/zero3_bf16.json

load_in_8bit: false
load_in_4bit: false

# CCE disabled (aarch64/torch2.9 grad explosion — see baselines/sera/README.md)
# plugins:
#   - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

# Use the base model's own chat_template (from Qwen/Qwen3-8B tokenizer_config.json)
# so training-time rendering matches vLLM's inference-time rendering byte-for-byte.
# The prior `chat_template: chatml` was a bare <|im_start|>role\ncontent<|im_end|>
# template that doesn't strip `<think>` blocks from prior assistant turns, while
# stock Qwen3-8B's template DOES strip them (and pads the last one with newlines).
# That mismatch between training and inference multi-turn contexts caused the
# model to go OOD after ~2 tool-call turns → whitespace collapse → 0% pass rate.
# See agent-traces-analysis/SMOKE_TEST_FINDING.md for the full diagnosis.
chat_template: tokenizer_default
datasets:
  - path: laion/Sera-4.6-Lite-T2-v4-316
    type: chat_template
    field_messages: messages
    ds_type: json
    message_field_training: train

dataset_prepared_path: /e/data1/datasets/playground/ot-baf/axolotl_dataset_cache/sera-v4-316
output_dir: /e/data1/datasets/playground/ot-baf/checkpoints/sera-v4-316-axolotl__Qwen3-8B

sequence_len: 32768

wandb_project:
wandb_entity:
wandb_watch:
wandb_name: sera-v4-316-axolotl__Qwen3-8B
wandb_log_model:

gradient_accumulation_steps: 8
micro_batch_size: 1
num_epochs: 3
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 1e-5
adam_beta1: 0.9
adam_beta2: 0.95

bf16: auto
tf32: false

gradient_checkpointing: true
activation_offloading: true
resume_from_checkpoint:
logging_steps: 1
flash_attention: true

loss_watchdog_threshold: 5.0
loss_watchdog_patience: 3

warmup_ratio: 0.1875
evals_per_epoch: 0
save_strategy: epoch

weight_decay: 0.01
max_grad_norm: 1.0
special_tokens:

e/data1/datasets/playground/ot-baf/checkpoints/sera-v4-316-axolotl__Qwen3-8B

This model is a fine-tuned version of Qwen/Qwen3-8B on the laion/Sera-4.6-Lite-T2-v4-316 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-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 8
• total_train_batch_size: 32
• total_eval_batch_size: 4
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.95) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 3
• training_steps: 17

Training results

Framework versions

• Transformers 5.5.0
• Pytorch 2.9.1+cu130
• Datasets 4.5.0
• Tokenizers 0.22.2