djdumpling/qwen3-4b-instruct-megagem-distill-para

TEXT GENERATIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 13, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The djdumpling/qwen3-4b-instruct-megagem-distill-para model is a 4 billion parameter instruction-tuned language model based on the Qwen3 architecture, specifically distilled for paraphrase generation. It was developed by djdumpling using a novel ExIt training method involving self-play games and an analytic expert, resulting in improved performance against Gemini-3-Flash. This model is optimized for generating paraphrased responses, with its training incorporating a self-paraphrasing step in the blueprint's own voice.

Loading preview...

Model Overview

The djdumpling/qwen3-4b-instruct-megagem-distill-para is a 4 billion parameter instruction-tuned model built upon the Qwen3 architecture. It integrates a paraphrase distillation LoRA (distill1a_d1a_para) with an SFT blueprint, making it ready for deployment with standard vLLM/transformers without additional components.

Unique Training Methodology

This model employs a distinctive training approach called ExIt with an analytic expert. This method involves:

  • Self-play games: The model engages in self-play to generate data.
  • Analytic expert: A fitted opponent bid-distribution price law (F̂) and a self-play value head (V̂) are derived from these games.
  • Closed-form best-response labels: These labels are generated on deviation turns based on expected surplus.
  • Self-paraphrasing: The blueprint's own voice is used to self-paraphrase these labels, which are then used for SFT LoRA.

Notably, this training process does not rely on RL gradients or external opponent-API data for any learned component; the price law is fit exclusively on the blueprint's own self-play bids.

Performance Highlights

When evaluated against 2x Gemini-3-Flash models (paired CRN, weights-only, without a test-time selector), this model achieved a significant improvement:

  • Win-rate: Increased from 29.3% to 40.7%.
  • Score/game: A CV-adjusted gain of +7.4 (t≈3.0).

For deployment, an additional zero-LLM-cost analytic EV selector can be integrated, which is external code and not part of these weights, further enhancing performance against Flash/Pro/Sonnet by approximately +11–14 points.