metadata
base_model: Qwen/Qwen3.5-9B
library_name: peft
model_name: output
tags:
- base_model:adapter:Qwen/Qwen3.5-9B
- lora
- sft
- transformers
- trl
licence: license
pipeline_tag: text-generation
output
This model is a fine-tuned version of Qwen/Qwen3.5-9B.
Training procedure
Hyperparameters
| Parameter | Value |
|---|---|
| Learning rate | 0.0001 |
| LR scheduler | SchedulerType.COSINE |
| Per-device batch size | 1 |
| Effective batch size | 1 |
| Epochs | 5 |
| Max sequence length | 2048 |
| Optimizer | OptimizerNames.PAGED_ADEMAMIX_8BIT |
| Weight decay | 0.01 |
| Warmup ratio | 0.05 |
| Max gradient norm | 1.0 |
| Precision | bf16 |
| Loss type | nll |
LoRA configuration
| Parameter | Value |
|---|---|
| Rank (r) | 32 |
| Alpha | 8 |
| Target modules | attn.proj, down_proj, gate_proj, in_proj_a, in_proj_b, in_proj_qkv, in_proj_z, k_proj, linear_fc1, linear_fc2, o_proj, out_proj, q_proj, qkv, up_proj, v_proj |
| rsLoRA | yes |
| Quantization | 4-bit (nf4) |
Dataset statistics
| Dataset | Samples | Total tokens | Trainable tokens |
|---|---|---|---|
| ken_instruct_250_cleaned.jsonl | 242 | 201,326 | 126,571 |
Training config
model_name_or_path: Qwen/Qwen3.5-9B
bf16: true
gradient_checkpointing: true
gradient_checkpointing_kwargs:
use_reentrant: false
use_liger: true
max_length: 2048
last_assistant_only_loss: true
learning_rate: 0.0001
warmup_ratio: 0.05
weight_decay: 0.01
lr_scheduler_type: cosine
neftune_noise_alpha: 5
aux_loss_top_prob_weight: 0.1
per_device_train_batch_size: 1
gradient_accumulation_steps: 1
optim: paged_ademamix_8bit
max_grad_norm: 1.0
use_peft: true
load_in_4bit: true
bnb_4bit_quant_type: nf4
lora_r: 32
lora_alpha: 8
lora_dropout: 0
use_rslora: true
logging_steps: 1
disable_tqdm: false
save_strategy: steps
save_steps: 500
save_total_limit: null
report_to: wandb
output_dir: output
data_config: data.yaml
prepared_dataset: prepared
num_train_epochs: 5
saves_per_epoch: 1
run_name: qwen35-9b-ken-v2
Data config
datasets:
- path: ken_instruct_250_cleaned.jsonl
type: conversational
truncation_strategy: drop
columns:
- messages
shuffle_datasets: true
shuffle_combined: true
shuffle_seed: 42
eval_split: 0.0
split_seed: 42
Framework versions
- PEFT 0.18.1
- Loft: 0.1.0
- Transformers: 5.2.0
- Pytorch: 2.6.0+cu124
- Datasets: 4.6.1
- Tokenizers: 0.22.2