README

README.md

@@ -3,39 +3,28 @@ license: apache-2.0
 language: [en]
 tags: [reinforcement-learning, data-science, code-agent, jupyter, harbor, curriculum, difficulty]
 size_categories: [1K<n<10K]
-configs:
-- config_name: default
-  data_files: "*.parquet"
 ---
 
 # AdithyaSK/data_agent_rl_environment_train_difficulty_ranked
 
-**Empirical difficulty ranking** of the 2238 tasks in
-[`AdithyaSK/data_agent_rl_environment_train`](https://huggingface.co/datasets/AdithyaSK/data_agent_rl_environment_train),
-ordered **easy → hard** (`rank` 1 = easiest).
+A **Harbor task suite** of 2238 data-agent tasks, **ordered easy -> hard** by *empirical*
+difficulty measured from a pass@4 rollout sweep (Qwen3.5-4B + 2B, bash harness).
 
-Derived from a **pass@4 rollout sweep** (Qwen3.5-4B + Qwen3.5-2B, bash harness, 4 attempts each,
-up to 8 attempts/task) — ranked from what the models *actually did*, not an LLM's guess.
+## Layout (standard Harbor spec)
+```
+tasks/<task_id>/{task.toml, instruction.md, environment/, tests/}
+registry.json     # tasks[] IN DIFFICULTY ORDER (rank 1 = easiest); each entry has rank/difficulty/solve_frac
+manifest.json     # full ranked table (all signals + task metadata)
+```
 
-## Columns
-| column | meaning |
-|---|---|
-| `rank` | 1 = easiest … 2238 = hardest |
-| `difficulty` | composite 0..1 (0 easiest). `0.70·(1-solve_frac) + 0.12·calls + 0.10·tokens + 0.08·time` |
-| `solve_frac` | fraction of all attempts that passed (primary signal) |
-| `pass4_4b`, `pass4_2b` | did the model solve it in ≥1 of its 4 attempts (0/1) |
-| `mean_calls` | avg bash tool calls per attempt |
-| `mean_elapsed` | avg wall-clock seconds per attempt |
-| `med_tokens` | median completion tokens per attempt |
-| `no_answer_frac` | fraction of attempts that never submitted |
-| `question`, `kaggle_dataset_name`, `reward_mode_initial` | from the task |
-| `difficulty_level` | the earlier **LLM-assigned** label (L1–L5) — for comparison |
+## The ordering
+`registry.json`'s `tasks[]` array is sorted easy -> hard. Each entry:
+`{name, path, rank, difficulty, solve_frac, mean_tool_calls}`.
 
-## Validation
-Empirical difficulty rises monotonically with the LLM-assigned `difficulty_level`
-(L1≈0.21 → L5≈0.84), confirming the ranking is sound while adding fine-grained ordering
-*within* each level (solve rate + tool calls + time + tokens).
+`difficulty` (0 easiest .. 1 hardest) = `0.70·(1-solve_frac) + 0.12·tool_calls + 0.10·tokens + 0.08·time`,
+from the real rollouts. Validated against the earlier LLM difficulty labels (monotonic L1->L5).
 
-## Use
-Curriculum RL (train easy→hard), difficulty-stratified eval, or filtering to the
-partial-solve band (`0 < solve_frac < 1`) which carries the most GRPO signal.
+## Tasks
+Same 2238 tasks as [`AdithyaSK/data_agent_rl_environment_train`](https://huggingface.co/datasets/AdithyaSK/data_agent_rl_environment_train),
+but each carries the **base image** (`savatar101/env-data-agent-train:base`) and the **3-reward
+verifier** (`correctness`, `submission`, `tool_efficiency` -> `reward.json`). Curriculum-ready.