---
library_name: transformers
base_model: Qwen/Qwen3-8B
tags:
- generated_from_trainer
# datasets: (stripped — invalid local path)
_datasets_:
- laion/Sera-4.6-Lite-T2-v4-1000
model-index:
- name: e/data1/datasets/playground/ot-baf/checkpoints/sera-v4-1000-axolotl__Qwen3-8B-v6
  results: []
---

<!-- 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"/>](https://github.com/axolotl-ai-cloud/axolotl)
<details><summary>See axolotl config</summary>

axolotl version: `0.16.0.dev0`
```yaml
# Sera v6 — scale data 316→1000 + num_epochs 3→6.
#
# Background: Sera v3 (316 rows × 6 epochs, SLURM 391242) passed turn-1 cleanly
# but collapsed at turn-3+ (degenerate tokens, 4.4.4.4… or for-the-for-the…)
# once a tool observation >~20 KB entered context. Greedy decoding didn't save
# it, so the root cause is under-training rather than sampling. See
# /Users/benjaminfeuer/Documents/notes/ot-agent/sera_braces_diagnosis.md for
# evidence (per-token probe + turn-3 replay).
#
# v6 = F3 fix: 3× more rows to give the model enough updates to stay stable
# in long multi-turn contexts.

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

chat_template: tokenizer_default
# datasets: (stripped — invalid local path)
_datasets_:
- laion/Sera-4.6-Lite-T2-v4-1000
    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-v4-1000-v6
output_dir: /e/data1/datasets/playground/ot-baf/checkpoints/sera-v4-1000-axolotl__Qwen3-8B-v6

sequence_len: 32768

wandb_project:
wandb_entity:
wandb_watch:
wandb_name: sera-v4-1000-axolotl__Qwen3-8B-v6
wandb_log_model:

gradient_accumulation_steps: 8
micro_batch_size: 1
num_epochs: 6
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:

```

</details><br>

# e/data1/datasets/playground/ot-baf/checkpoints/sera-v4-1000-axolotl__Qwen3-8B-v6

- laion/Sera-4.6-Lite-T2-v4-1000

## 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: 20
- training_steps: 109

### Training results



### Framework versions

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