L3-8B-Stheno-v3.3-32K-8.0bpw-h8-exl2

Trained with compute from Backyard.ai | Thanks to them and @dynafire for helping me out.

---

Exl2 quant done of Stheno 3.3 32k by Sao10k

Training Details: <br>Trained at 8K Context -> Expanded to 32K Context with PoSE training.

Dataset Modifications: <br>- Further Cleaned up Roleplaying Samples -> Quality Check <br>- Removed Low Quality Samples from Manual Check -> Increased Baseline Quality Floor <br>- More Creative Writing Samples -> 2x Samples <br>- Remade and Refined Detailed Instruct Data

Notes: <br>- Training run is much less aggressive than previous Stheno versions. <br>- This model works when tested in bf16 with the same configs as within the file. <br>- I do not know the effects quantisation has on it. <br>- Roleplays pretty well. Feels nice in my opinion. <br>- It has some issues on long context understanding and reasoning. Much better vs rope scaling normally though, so that is a plus. <br>- Reminder, this isn't a native 32K model. It has it's issues, but it's coherent and working well.

Sanity Check // Needle in a Haystack Results: <br>- This is not as complex as RULER or NIAN, but it's a basic evaluator. Some improper train examples had Haystack scores ranging from Red to Orange for most of the extended contexts.

Wandb Run:

---

Relevant Axolotl Configurations: <br>-> Taken from winglian/Llama-3-8b-64k-PoSE <br>- I tried to find my own configs, hours of tinkering but the one he used worked best, so I stuck to it. <br>- 2M Rope Theta had the best loss results during training compared to other values. <br>- Leaving it at 500K rope wasn't that much worse, but 4M and 8M Theta made the grad_norm values worsen even if loss drops fast. <br>- Mixing in Pretraining Data was a PITA. Made it a lot worse with formatting. <br>- Pretraining / Noise made it worse at Haystack too? It wasn't all Green, Mainly Oranges. <br>- Improper / Bad Rope Theta shows in Grad_Norm exploding to thousands. It'll drop to low values alright, but it's a scary fast drop even with gradient clipping.

sequence_len: 8192
use_pose: true
pose_max_context_len: 32768

overrides_of_model_config:
  rope_theta: 2000000.0
  max_position_embeddings: 32768

  # peft_use_dora: true
adapter: lora
peft_use_rslora: true
lora_model_dir:
lora_r: 256
lora_alpha: 256
lora_dropout: 0.1
lora_target_linear: true
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj

warmup_steps: 80
gradient_accumulation_steps: 6
micro_batch_size: 1
num_epochs: 2
optimizer: adamw_bnb_8bit
lr_scheduler: cosine_with_min_lr
learning_rate: 0.00004
lr_scheduler_kwargs:
    min_lr: 0.000004