Andy-ML-And-AI's picture
Update README.md
bc48c43 verified
metadata
base_model: unsloth/qwen3-8b-unsloth-bnb-4bit
tags:
  - text-generation-inference
  - transformers
  - unsloth
  - qwen3
  - trl
  - qlora
  - reasoning
  - code
  - hyperthinkcode
license: apache-2.0
language:
  - en
datasets:
  - Sashvat/HyperThink-X-Nvidia-Opencode-Reasoning-200K
metrics:
  - humaneval
  - gsm8k
library_name: adapter
pipeline_tag: text-generation

HyperThinkCode-Qwen3-8B-v1

HyperThinkCode-Qwen3-8B-v1 is a LoRA fine-tune of the Qwen3-8B base model.


πŸ›  Experimental Setup

  • Base model: Qwen3-8B
  • Hardware: dual Tesla T4 (16GB VRAM each)
  • 4-bit QLoRA with rank = 16 and alpha = 16
  • All linear layers:
    • Attention: q, k, v, o
    • MLP: gate, up, down
  • Training time: ~1 hour 17 minutes
  • Total steps: 50

🧠 Dataset & Objective

Training on a specific 30k subset of the
Sashvat/HyperThink-X-Nvidia-Opencode-Reasoning-200K dataset.

  • Uses chat template with assistant response in the thinking field
  • Objective: encourage thinking over direct response
  • Sequence length limited to 4096 tokens (for code complexity + VRAM constraints)

πŸ“‰ Training Logs

With only 50 steps, the loss shows expected variance given model + dataset complexity.

Step Training Loss
10 0.8177
25 0.7358
50 0.6785
  • Global batch size: 8 (1 device Γ— 8 gradient steps)

πŸ“Š Evaluation (Ongoing)

Currently running benchmarks using the lm-eval library:

  • HumanEval (Coding)
  • GSM8K (Math)

Comparisons are being made against the base model.


πŸ” Reproduction

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "Andy-ML-And-AI/HyperThinkCode-Qwen3-8B-v1",
    max_seq_length = 4096,
    load_in_4bit = True,
)