Qwen3.5-4B-Winner-SFT-Epochs

Qwen3.5-4B Winner-Only Debate SFT

This public repository contains full-parameter checkpoints from three epochs of supervised fine-tuning of Qwen/Qwen3.5-4B on judge-selected winning arguments from multi-round, Wikipedia-grounded debates:

• epoch-1-step-315/
• epoch-2-step-630/
• epoch-3-step-945/

The exact epoch-3 checkpoint used in the eight-model debate evaluation has model.safetensors SHA-256 9160631bf9cbcd66bb3e0b18c074a5a343dadfc6d2d9616d53f80088025ca0c9.

The original rolling-checkpoint run retained epoch 3 but pruned the exact epoch-1 and epoch-2 files. Epochs 1 and 2 were therefore reconstructed from the same data, configuration, seeds, software, and three-GPU setup, saving model weights at each epoch boundary. provenance.json records every weight hash and whether the reproduced epoch-3 checkpoint is byte-identical to the evaluated epoch-3 checkpoint. If it differs, the reproduced epoch-3 weights are also retained under reproduced-run/epoch-3-step-945/.

Training

• Training examples: 3,780 conversations from 180 topics
• Validation examples: 420 conversations from 20 disjoint topics
• Supervision: all four arguments from each judge-selected winning side
• Fine-tuning: full language-model weights; not LoRA
• Context cutoff: 4,096 tokens
• Epochs: 3
• Effective batch size: 12
• Learning rate: 1e-5
• Scheduler: cosine with 5% warmup
• Precision: BF16
• Distributed training: 3 GPUs with DeepSpeed ZeRO-3
• Random seed and data seed: 42
• Prompt template: qwen3_5_nothink

Evaluation

The exact epoch-3 model debated seven opponents on the same 200 topics used to derive its training conversations. An identity-blind OpenAI GPT-5.6-Sol evaluator selected one winner per debate.

• Overall: 718 wins / 682 losses (51.29%), rank 5 of 8
• gemma-e2b: 118 / 200 (59.0%)
• gemma-e4b: 116 / 200 (58.0%)
• base Qwen3.5-4B: 113 / 200 (56.5%)
• Qwen3.5-9B: 97 / 200 (48.5%)
• Qwen3.5-27B: 92 / 200 (46.0%)
• Qwen3.5-122B-A10B-FP8: 92 / 200 (46.0%)
• Qwen3.5-2B: 90 / 200 (45.0%)

This is an in-sample, training-topic evaluation and must not be interpreted as held-out generalization.

Loading a checkpoint

from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "Sarim-Hash/Qwen3.5-4B-Winner-SFT-Epochs"
subfolder = "epoch-3-step-945"

tokenizer = AutoTokenizer.from_pretrained(repo_id, subfolder=subfolder)
model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    subfolder=subfolder,
    torch_dtype="auto",
    device_map="auto",
)

Use the supplied chat template with thinking disabled to match training and evaluation.