SaylorTwift HF Staff
eval: pass purpose=eval (server requires it with eval_sampling)
e7a4c0e verified | """Held-out evaluation: base model vs LoRA adapter on SmolDataEnvs test tasks. | |
| Runs N test tasks through the env server under OpenCode, once per model, and | |
| reports pass@1 (first graded attempt correct) for each, plus the per-task | |
| agreement breakdown the article's comparisons used. | |
| The vLLM server must serve the base model AND the adapter as a named LoRA: | |
| vllm serve LiquidAI/LFM2.5-2.6B --enable-lora --max-lora-rank 32 \ | |
| --lora-modules trained=<adapter-repo> ... | |
| Config via env: | |
| ENV_URL env server (default http://127.0.0.1:8000) | |
| VLLM_URL policy server (default http://127.0.0.1:8001) | |
| TEST_SPLIT served test split (default /app/prepared/datasets/test) | |
| BASE_MODEL served base model id (default LiquidAI/LFM2.5-2.6B) | |
| ADAPTER_NAME LoRA name on the server (default "trained"); "" evaluates base only | |
| N_TASKS how many test tasks (default 50) | |
| CONCURRENCY parallel rollouts (default 8; keep <= MAX_CONCURRENT_ENVS) | |
| OUT results JSON path (default /app/eval_results.json) | |
| """ | |
| import json | |
| import os | |
| import threading | |
| from concurrent.futures import ThreadPoolExecutor, as_completed | |
| from harbor_env import HarborEnv # re-export of openenv.harbor.client.HarborEnv | |
| ENV_URL = os.environ.get("ENV_URL", "http://127.0.0.1:8000") | |
| VLLM_URL = os.environ.get("VLLM_URL", "http://127.0.0.1:8001") | |
| TEST_SPLIT = os.environ.get("TEST_SPLIT", "/app/prepared/datasets/test") | |
| BASE_MODEL = os.environ.get("BASE_MODEL", "LiquidAI/LFM2.5-2.6B") | |
| ADAPTER_NAME = os.environ.get("ADAPTER_NAME", "trained") | |
| N_TASKS = int(os.environ.get("N_TASKS", "50")) | |
| CONCURRENCY = int(os.environ.get("CONCURRENCY", "8")) | |
| OUT = os.environ.get("OUT", "/app/eval_results.json") | |
| _local = threading.local() | |
| def run_one(task_index, model): | |
| """One graded rollout. Returns (correct, failed, error_or_none).""" | |
| if not hasattr(_local, "env"): | |
| _local.env = HarborEnv(base_url=ENV_URL) | |
| kwargs = dict( | |
| split=TEST_SPLIT, | |
| task_index=task_index, | |
| harness="opencode", | |
| sandbox=os.environ.get("SANDBOX", "hf-sandbox"), | |
| reward_key="correctness,reward", | |
| llm_url=VLLM_URL, | |
| model=model, | |
| purpose="eval", # required by the server when eval_sampling is passed | |
| eval_sampling={"temperature": 1.0, "top_p": 1.0}, | |
| agent_step_limit=17, # production eval cap from the FineEnvs Space | |
| agent_timeout_sec=600, | |
| ) | |
| try: | |
| result = _local.env.run_rollout(**kwargs) | |
| except TypeError: | |
| # older server without eval_sampling: retry without it | |
| kwargs.pop("eval_sampling") | |
| result = _local.env.run_rollout(**kwargs) | |
| except Exception as exc: | |
| return task_index, model, None, True, f"{type(exc).__name__}: {exc}" | |
| if not result.ok or result.reward is None: | |
| return task_index, model, None, True, getattr(result, "error", None) or "not ok" | |
| return task_index, model, float(result.reward) == 1.0, False, None | |
| def main(): | |
| models = [BASE_MODEL] + ([ADAPTER_NAME] if ADAPTER_NAME else []) | |
| work = [(i, m) for i in range(N_TASKS) for m in models] | |
| cells = {} # (task_index, model) -> {"correct": bool|None, "failed": bool, "error": str|None} | |
| with ThreadPoolExecutor(max_workers=CONCURRENCY) as pool: | |
| futures = [pool.submit(run_one, i, m) for i, m in work] | |
| done = 0 | |
| for fut in as_completed(futures): | |
| i, m, correct, failed, err = fut.result() | |
| cells[(i, m)] = {"correct": correct, "failed": failed, "error": err} | |
| done += 1 | |
| if done % 20 == 0 or done == len(work): | |
| print(f"[eval] {done}/{len(work)} cells done", flush=True) | |
| summary = {"n_tasks": N_TASKS, "models": {}} | |
| for m in models: | |
| scorable = [c["correct"] for (i, mm), c in cells.items() if mm == m and not c["failed"]] | |
| failed = sum(1 for (i, mm), c in cells.items() if mm == m and c["failed"]) | |
| summary["models"][m] = { | |
| "pass_at_1": sum(scorable) / len(scorable) if scorable else None, | |
| "solved": sum(scorable), | |
| "scorable": len(scorable), | |
| "failed_rollouts": failed, | |
| } | |
| if len(models) == 2: | |
| b, a = models | |
| both = sum(1 for i in range(N_TASKS) | |
| if cells.get((i, b), {}).get("correct") and cells.get((i, a), {}).get("correct")) | |
| only_a = sum(1 for i in range(N_TASKS) | |
| if not cells.get((i, b), {}).get("correct") and cells.get((i, a), {}).get("correct")) | |
| only_b = sum(1 for i in range(N_TASKS) | |
| if cells.get((i, b), {}).get("correct") and not cells.get((i, a), {}).get("correct")) | |
| summary["agreement"] = {"both_solved": both, "only_adapter": only_a, "only_base": only_b} | |
| pa = summary["models"][a]["pass_at_1"] | |
| pb = summary["models"][b]["pass_at_1"] | |
| if pa is not None and pb is not None: | |
| summary["delta"] = pa - pb | |
| print(json.dumps(summary, indent=2)) | |
| with open(OUT, "w") as f: | |
| json.dump({"summary": summary, | |
| "cells": {f"{i}|{m}": c for (i, m), c in cells.items()}}, f, indent=2) | |
| print(f"[eval] wrote {OUT}") | |
| if __name__ == "__main__": | |
| main() | |