DJLougen/eljefe-router-data

ElJefe Router Data — Counterfactual Local-vs-Frontier Pairs

Every row is a prompt answered by both a local model (google/gemma-4-E4B-it, bf16 on an L4) and a frontier model, with both answers graded deterministically — exact-match, numeric equivalence, code-test execution, constraint checking. No LLM judge. This is the training data for DJLougen/eljefe-v0.

What "counterfactual" means here

For each prompt we know what would have happened under both routing decisions. That's what makes this a routing dataset rather than a benchmark: you can ask "was escalating worth it?" per row, not just "did the model answer well?"

Files

File	Rows	Contents
tasks.jsonl	12,484	Normalized prompts (gsm8k, math500, mmlu_pro, mbpp, ifeval) with gold references
scored.jsonl	12,475	Full counterfactuals: E4B + deepseek-v4-flash-0731 answers, both graded
router_dataset.parquet	12,475	Training-ready rows with splits and derived labels
scored_glm5p3flash.jsonl	997	Same prompts × glm-5p3-flash frontier (1k subset)
scored_gptoss120b.jsonl	997	Same prompts × gpt-oss-120b frontier (1k subset)
router_dataset_{glm5p3flash,gptoss120b}.parquet	997	Router datasets for the alternate frontiers
mac_calibration.jsonl	200	Held-out calibration set for a specific local build

Schema (scored.jsonl)

{
  "id": "...", "source": "gsm8k", "task_family": "math",
  "prompt": "...", "system_prompt": null,
  "local_model": "google/gemma-4-E4B-it",
  "frontier_model": "accounts/fireworks/models/deepseek-v4-flash-0731",
  "local_answer": "...", "frontier_answer": "...",
  "local_score": 1.0, "frontier_score": 1.0, "delta_q": 0.0,
  "local_input_tokens": 540, "local_output_tokens": 220,
  "frontier_input_tokens": 540, "frontier_output_tokens": 181,
  "frontier_cost": 0.0004, "grader_type": "deterministic",
  "grader_confidence": 1.0
}

Derived labels (in the parquet): local_sufficient = local_score >= 0.70 AND delta_q <= 0.10; frontier_helpful = delta_q > 0.10.

When we say oracle anywhere in this project, we mean a cheat-mode router computed after the fact: since both models ran on every prompt, we know each row's actual local_score and frontier_score, and the oracle picks the better route per row (local_score >= frontier_score → local). It's not deployable — it "knows" the answer before choosing — but it marks the ceiling: the best possible cost-quality tradeoff any router could achieve on this data.

Splits

router_dataset.parquet carries group-aware splits (near-duplicate prompts stay together):

Split	Rows	Purpose
train	9,553	fit routers
validation	1,195	threshold + calibration fitting
test_iid	995	held-out same-distribution eval
test_ood	532	held-out task family (ifeval)
mac_calibration	200	per-device local-build calibration

Loading

import pandas as pd
df = pd.read_parquet("hf://datasets/DJLougen/eljefe-router-data/router_dataset.parquet")
# metadata column is JSON-serialized; decode if needed:
import json
df["metadata"] = df["metadata"].map(lambda s: json.loads(s) if isinstance(s, str) else s)

Things you can do with it

• Train your own router on any local/frontier pair — the schema is model-agnostic.
• Study the boundary: which task families need frontier (reasoning dominates), where local is enough, where frontier actually hurts (instruction-following).
• Frontier-swap analysis: the two alternate-frontier files let you test whether a routing boundary learned against one frontier transfers to another (it mostly does — ~95% of oracle utility).
• Calibration: mac_calibration.jsonl is a fixed 200-row set for measuring a specific local build (e.g. your quantized MLX/GGUF E4B) rather than the canonical checkpoint.

Reproduce

Full pipeline (fetch → generate → grade → split → train → evaluate): https://github.com/DJLougen/ElJefe — see README and PLAN_ElJefe_Colab_CLI.md.