Taggert2222

gpt-oss-20b-antidoom

README

License: apache-2.0

Results

These are our own evals on one box, October 2026, with stock and this model served by the same vLLM 0.31 (XPU) build and flags. The sample sizes are small, so read single-digit differences as noise.

Table with columns: stock gpt-oss-20b, this model
	stock gpt-oss-20b	this model
repetition loops, 200 held-out prompts, effort medium	17/200 (8.5%)	0/200
repetition loops, same 200, effort high	79/200 (39.5%)	2/200 (1.0%)
graded accuracy, 700 items (GSM8K, MATH-500, MMLU, CommonsenseQA), medium	78.4%	83.0%
code inside a tool call (PTC probe), 72 calls, medium	57/72	57/72
multi-turn streaming tool sessions (2 × 16), format faults	0	0
coding-agent tasks in our agent harness, 28 each, verified on disk	17 pass	20 pass

Graded eval: paired on the same 700 items, this model fixes 47 and breaks 15 (McNemar p < 0.001). Many of the fixes are items where stock looped or ran out of tokens.
Soak: this model has served as our live gpt-oss for coding agents since October 17, 2026. A 4-hour, 16-stream soak answered 5,180 requests with one proxy error and no server errors, and about 0.1% of its generations looped.
Speed: unchanged. The tensors and shapes are identical to stock, and its replies are shorter: median 305 completion tokens on the loop set at medium, against 418 for stock.

Use

It's a drop-in replacement for openai/gpt-oss-20b, with the same harmony chat format, tokenizer and chat template.

vllm serve <this repo> --served-model-name gpt-oss --tool-call-parser openai --enable-auto-tool-choice \  --reasoning-parser openai_gptoss

Sampling: OpenAI's recommendation for gpt-oss, temperature 1.0 and top_p 1.0.
Reasoning effort: medium is the best balance. At high effort, stock gpt-oss-20b loops often. This model rarely does, but it still reasons long, so give it a generous max_tokens.
Tool calls: for a strict tool-call grammar and an effort cap on vLLM 0.31, see the cookbook's gptoss031 plugin.

What changed in the weights

LoRA: rank 256, alpha 128, on attention q_proj, k_proj, v_proj and o_proj in all 24 layers. The learning rate was 1.5e-5, with early stop at step 25.
Merge: we merged it as W + (alpha/r)·B·A, in fp32, into the bf16 attention tensors: 96 tensors. The MXFP4 experts, embeddings, router and lm_head are the original bytes.
Training data: 2,959 preference pairs. Each pair is a loop-start token (rejected) against the model's own top-logprob alternatives (chosen). They were mined from antidoom-mix-v1.0 prompts at medium and high effort, including tool-calling coding tasks rendered with the exact serving chat template.
Training hardware: two Intel Arc Pro B70s, bf16, eager attention, max sequence length 5,120.

The unmerged adapter is in adapter/. Its target is the bf16 attention weights of openai/gpt-oss-20b. To apply it to the MXFP4 checkpoint yourself, use the cookbook's antidoom/merge_into_mxfp4.py.

Limitations

Small evals: all numbers come from our own evals, run once, on Intel XPU with vLLM 0.31. We did not test other GPUs or engines.
Narrated tool calls: like stock, inside a long agent prompt the model sometimes writes a tool call or an invented result as text instead of calling the tool. Our tests point to two agent-harness messages as the trigger: a mid-turn user-role status note, and a tool result that sounds like the task is finished. Changing those two messages fixed most of it. Training against it broke other behaviour. The cookbook's docs/AGENT-HARNESS.md covers this.
Same base: knowledge, safety behaviour and everything else are stock gpt-oss-20b's. OpenAI's gpt-oss usage policy applies.

Credits

Liquid AI: Antidoom (Apache-2.0), the loop detector, the pair mining, the FTPO trainer and the antidoom-mix prompt set. It builds on Antislop by Sam Paech et al.
OpenAI: gpt-oss-20b (Apache-2.0) and the harmony format.
Claude (Anthropic): ported the trainer to Intel XPU, ran the mining, training and evaluations, and wrote this card, through Claude Code.

License

Apache 2.0, the same as the base model.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

Taggert2222

Model Tree

Base

openai/gpt-oss-20b

Fine-tuned

this model

Input Modalities

Text

Output Modalities