Qwen3.5-Reverse-Text-SFT

A short SFT fine-tune of Qwen/Qwen3.5-0.8B on PrimeIntellect/Reverse-Text-SFT. It is a small, deliberately under-trained starting point for reverse-text RL, made for the planned move of the prime-rl CI RL tests from PrimeIntellect/Qwen3-0.6B-Reverse-Text-SFT to Qwen3.5. It is not meant for general use.

Recipe

• Code: prime-rl commit 21814b401 (branch ci/qwen3_5-ci, contains the Qwen3.5 tied lm_head fix #3863 and the CP fix #3864).
• Config (uv run sft @ sft.toml), 1 H200:

max_steps = 10

[model]
name = "Qwen/Qwen3.5-0.8B"

[data]
name = "PrimeIntellect/Reverse-Text-SFT"
seq_len = 4096
batch_size = 32

[optim]
lr = 2e-5

• Chat template: unchanged Qwen3.5 template, thinking off (the 0.8B default). Completions are rendered with the empty \n\n\n\n prefix, the same as RL generation.
• Weights include the (frozen, unchanged) vision tower, so the checkpoint loads as Qwen3_5ForConditionalGeneration like the base model.

Numbers

• SFT loss, steps 1-10: 4.98, 6.16, 5.59, 4.96, 4.53, 4.22, 3.96, 3.62, 3.27, 2.92.
• reverse-text eval reward (LCS ratio, 256 prompts, temperature 1, 128 max tokens): 0.31 (base model: 0.03 on 32 prompts).