yvligo

yvligo is DeepSeek-V4.1-Flash-1017 post-trained to run X (Twitter) Ads campaigns as a tool-using agent. Given a client's brief, it:

• sets up campaigns through X Ads API tools;
• checks on them through the flight and adjusts targeting, bids and budgets;
• ends with a recap to the client.

It is a LoRA adapter trained with supervised fine-tuning on expert traces: complete campaign runs by an expert agent in a simulated X Ads environment. A run became a trace only when it beat a scripted baseline and broke none of the brief's rules.

Results

We ran each agent once on the 45 briefs of AdVal, our benchmark of agents running ad campaigns on a simulated X Ads market, at the same rollout seed for every agent. None of these briefs or their worlds appear in the training data.

A run scores its conversions over par: the expected conversions its campaigns drove, divided by those of a scripted baseline in the same world and seed. The baseline runs one ad group on the brief's rules at X's default AUTO bid.

Agent	Geometric mean of conversions ÷ par	yvligo ÷ agent, paired by brief (SE)	Briefs yvligo is ahead on
Expert (source of the training traces)	1.349	×0.972 (×1.025)	47%
yvligo	1.311
Gemini 4 Argon (Antigravity)	1.198	×1.094 (×1.034)	67%
GPT-6.1 Sol (Codex)	1.145	×1.145 (×1.035)	73%
DeepSeek-V4.1-Flash, before training (same setup as yvligo)	1.095	×1.196 (×1.036)	80%
Grok 4.7 (Grok Build)	1.087	×1.206 (×1.038)	80%
Claude Opus 5.5 (Claude Code)	1.068	×1.228 (×1.033)	89%
GPT-6 Astra (Codex)	1.058	×1.239 (×1.036)	82%
Muse Spark 1.3 (Muse Code)	1.041	×1.259 (×1.040)	89%
Inkling 2 Small (Claude Code on Tinker)	1.023	×1.282 (×1.032)	91%
Claude Haiku 5.5 (Claude Code)	1.012	×1.295 (×1.037)	89%
Muse Wonder 1.0 (Muse Code)	1.007	×1.302 (×1.040)	89%
Claude Sonnet 5.5 (Claude Code)	0.930	×1.409 (×1.043)	96%
GPT-6 Luna (Codex)	0.738	×1.775 (×1.244)	93%
Nemotron 3.5 Lightning (Claude Code on OpenRouter)	0.213	×6.147 (×1.469)	96%

The other models ran in their own agent CLIs at medium reasoning effort.

Across the 45 briefs, yvligo:

• never scored under 0.8 of par (its lowest was 0.842);
• wrote an accurate recap every time;
• had 2.3% of its tool calls return an error (the expert 1.6%);
• broke one rule, as did the expert: on a phones-only brief it also advertised on desktop;
• gated to the brief's minimum age on 5 of the 11 briefs that set one; the expert never did.

X's policy and the brief set different kinds of requirements. X's policy is legal: alcohol ads in the US must target 21+, and X disapproves an ad group that breaks it. The brief says who can convert, such as UK residents on phones aged 18+; reaching anyone else wastes spend but breaks no law. Matching the brief's markets and platforms costs nothing, so missing them is an error. Matching its minimum age does cost: X only age-targets people whose age it knows, so an 18+ gate also drops adults of unknown age. Leaving it out can pay, and neither choice counts against a run.

Training

Traces. The expert ran 444 campaigns: 222 briefs, two rollout seeds each. A run was kept when it:

• finished;
• scored at least par;
• scored at least as well as the base model on the same brief and seed;
• passed the brief's rule checks.

That gave 312 traces from 181 of the 196 training briefs; the other 26 briefs were held back for validation. Each trace is a whole run of up to 262,144 tokens: the brief, every reasoning step, tool call and tool result, and the final recap.

Fine-tuning. Supervised fine-tuning on Tinker:

• loss on every assistant turn;
• LoRA rank 32, alpha 32, on all linear layers;
• Adam (β₁ 0.9, β₂ 0.95, ε 1e-8), 8 traces per step;
• learning rate 1e-4, decaying linearly over a planned 117 steps (three epochs).

Checkpoint. We checked the agent's behaviour after each training block (steps 10, 20, 39 and 59). This adapter is step 39: one epoch, about 24M tokens. The next checkpoint, at step 59, broke a tool-call format and was rejected. Validation loss on 22 held-out traces fell from 0.415 to 0.366.

Chat format. DeepSeek V4.1 with thinking at low effort, keeping earlier turns' reasoning in the context. We used tinker-cookbook's renderer, which matches DeepSeek's API token for token.

Source checkpoint: tinker://cd80698b-c69b-5061-b324-a416cee754e4:train:0/weights/block3-step39

How to use

This is a standard PEFT LoRA adapter for deepseek-ai/DeepSeek-V4.1-Flash-1017. Load it with any stack that serves the base model with LoRA adapters. With Transformers and PEFT:

from transformers import AutoModelForCausalLM
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("deepseek-ai/DeepSeek-V4.1-Flash-1017", torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(base, "loglogai/yvligo")

It was trained and measured in one agent setup:

• the system prompt is DeepSeek Harness's headless agent prompt;
• the user message is the client's brief;
• the tools are an X Ads API surface as MCP tools: 78 tools under mcp__ads__*, plus mcp__scheduler__wait_until to pause until later in the flight;
• thinking is on at low effort, with earlier turns' reasoning kept in the context.

It expects to act through tools and to end the session with a recap to the client.

Limitations

• Keyword gains. Part of its gain over the base model comes from keyword targeting, which the simulator may reward more than X does.
• Evaluation awareness. In 18 of the 45 runs, its reasoning remarks on being evaluated.

License

MIT, the same as the base model.