| import torch |
|
|
| paths = { |
| 0: "outputs_v2/Llama-3.2-3B-Instruct-MATH-500-tokens10-lr0.03-sigma0.1-sigdecay0.99-steps0-all-rewardall_tokens-conf/logistics.pt", |
| 1: "outputs_v2/Llama-3.2-3B-Instruct-MATH-500-tokens10-lr0.03-sigma0.1-sigdecay0.99-steps1-all-rewardall_tokens-conf/logistics.pt", |
| 2: "outputs_v2/Llama-3.2-3B-Instruct-MATH-500-tokens10-lr0.03-sigma0.1-sigdecay0.99-steps2-all-rewardall_tokens-conf/logistics.pt", |
| 4: "outputs_v2/Llama-3.2-3B-Instruct-MATH-500-tokens10-lr0.03-sigma0.1-sigdecay0.99-steps4-all-rewardall_tokens-conf/logistics.pt", |
| 10: "outputs_v2/Llama-3.2-3B-Instruct-MATH-500-tokens10-lr0.03-sigma0.1-sigdecay0.99-steps10-all-rewardall_tokens-conf/logistics.pt", |
| } |
|
|
| results = {} |
| for step, path in paths.items(): |
| ckpt = torch.load(path) |
| step_dict = {} |
| for entry in ckpt["entries"]: |
| step_dict[entry["data_idx"]] = { |
| "is_correct": entry["is_correct"], |
| "answer": entry["answer"], |
| "init_reward": entry.get("init_reward", None), |
| "best_reward": entry["best_reward"], |
| "best_reward_step": entry["best_reward_step"], |
| } |
| results[step] = step_dict |
|
|
| all_ids = sorted(set().union(*[set(v.keys()) for v in results.values()])) |
|
|
| for idx in all_ids: |
| row = [f"idx={idx}"] |
| init_r = results[0].get(idx, {}).get("init_reward", None) |
| row.append(f"init_r={init_r}") |
| row.append(f"s10_best_step={results[10].get(idx, {}).get('best_reward_step', None)}") |
| for step in [0, 1, 2, 4, 10]: |
| ok = results[step].get(idx, {}).get("is_correct", None) |
| row.append(f"s{step}={ok}") |
| print(" | ".join(row)) |
