agurung/flawed-fictions-qwen3-4b-litereason-sft-positive

Flawed Fictions Qwen3-4B LiteReason SFT (positive)

Projector SFT checkpoint trained on the positive-only Flawed Fictions implicit-thought dataset. This upload preserves the full base LM weights plus LiteReason reasoning_projector weights in the safetensors shards.

Training Details

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Dataset variant: positive (gold-label Yes subset only)
• Dataset/prep label in local artifacts: twenty_five_percent_masked
• W&B run: gm8xrkg2
• Source checkpoint directory: /mnt/disk/litereason_anon/litereason/experiments/flawed_fictions/outputs/sft_qwen3_4b_positive

What Is Saved

• Full Hugging Face model weights
• LiteReason reasoning_projector.* weights embedded in the uploaded safetensors
• config.json with is_litereason_model=true, litereason_num_reasoning_layers=3, and litereason_max_reasoning_steps=5
• Tokenizer / chat template files needed for reload

Usage

from litereason.causal_lm_with_reasoning import AutoModelForCausalLMWithReasoning

model = AutoModelForCausalLMWithReasoning.from_pretrained(
    "agurung/flawed-fictions-qwen3-4b-litereason-sft-positive",
    device_map="auto",
    torch_dtype="auto",
)