Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-step50
ttt-discover-v3-qwen3-8b-ac1-step50 is a machine learning model from EdwardoSunny. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
LoRA adapter (rank 32, all-linear) from the upstream-faithful local reproduction of TTT-Discover on autoresearch-bench ac1. Checkpoint saved after training step 49 (0-indexed). Strict upstream eval parity: 1100s hard…
Downloads · 30 days
0
Access
Public
Updated Jul 12, 2026
Repo size
349 MB
Likes
0
Public
Click a slice to open those files.
.safetensors349 MB · 100%
From the Hugging Face model README
LoRA adapter (rank 32, all-linear) from the upstream-faithful local
reproduction of TTT-Discover on autoresearch-bench ac1. Checkpoint saved
after training step 49 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 49,
"progress/batch": 49,
"optim/lr": 4e-05,
"progress/done_frac": 1.0,
"puct/buffer_size": 789,
"puct/sampled_size": 8,
"puct/T": 25088,
"puct/scale_last": 251.1762954598411,
"puct/buffer_value/mean": -1.8558982354612616,
"puct/buffer_value/std": 8.936086409855651,
"puct/buffer_value/min": -252.6826286848707,
"puct/buffer_value/max": -1.5063332250296204,
"puct/buffer_timestep/mean": 23.662864385297844,
"puct/buffer_timestep/std": 14.251601289813506,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 48.0,
"puct/buffer_construction_len/mean": 1055.1723700887198,
"puct/buffer_construction_len/std": 478.9746687244332,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7397.0,
"puct/sampled_value/mean": -1.5063332250296215,
"puct/sampled_value/std": 8.270975252328749e-16,
"puct/sampled_value/min": -1.5063332250296229,
"puct/sampled_value/max": -1.5063332250296204,
"puct/sampled_timestep/mean": 47.75,
"puct/sampled_timestep/std": 0.6614378277661477,
"puct/sampled_timestep/min": 46.0,
"puct/sampled_timestep/max": 48.0,
"puct/sampled_construction_len/mean": 1000.0,
"puct/sampled_construction_len/std": 0.0,
"puct/sampled_construction_len/min": 1000.0,
"puct/sampled_construction_len/max": 1000.0,
"time/sampling": 2560.0017008781433,
"env/all/ac_tokens_per_turn": 8439.294921875,
"env/all/ob_tokens_per_turn": 2895.375,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4320919,
"env/all/total_ob_tokens": 1482432,
"env/all/time/sampling_mean": 284.8152926629409,
"env/all/time/sampling_max": 387.4623713493347,
"env/all/time/env_step_mean": 821.9814522163942,
"env/all/time/env_step_max": 2194.5426433086395,
"env/all/reward/mean": 0.5822623980553665,
"env/all/reward/max": 0.663863729980263,
"env/all/reward/min": 0.0,
"env/all/format": 1.0,
"env/all/format/min": 1.0,
"env/all/format/max": 1.0,
"env/all/reward": 0.5822623980553665,
"env/all/correctness": 0.89453125,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 1.5497482463115166,
"env/all/raw_score/min": 1.5063332250296204,
"env/all/raw_score/max": 4.979628791308284,
"env/all/initial_raw_score": -1.5063332250296213,
"env/all/initial_raw_score/min": -1.5063332250296229,
"env/all/initial_raw_score/max": -1.5063332250296204,
"env/all/msg": "Success; raw_score=1.5063332250296204",
"env/all/parsed_code": "```python\nimport numpy as np\nimport cvxpy as cp\nimport time\n\ndef propose_candidate(seed=42, **kwargs):\n \"\"\"\n Proposes a sequence of non-negative numbers to minimize the evaluation score\n using an enhanced LP optimization approach with randomized perturbations,\n expanded constraint sets, and adaptive blending.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + 1000 - 10 # 1000 seconds with 10s buffer\n\n # Start with a safe, non-zero initial sequence (could be a better one)\n if 'height_sequence_1' in globals() and isinstance(height_sequence_1, (list, np.ndarray)):\n current_seq = np.array(list(height_sequence_1))\n else:\n n = 1000\n current_seq = np.full(n, 0.01)\n best_seq = current_seq.copy()\n best_score = evaluate_sequence(best_seq.tolist())\n n = len(current_seq)\n\n max_iterations = 300 # Increased for better exploration\n iteration = 0\n K = 10 # Number of top convolution positions to add as constraints\n t_decay_factor = 0.995 # Gradual decrease of blending parameter\n t_initial = 0.15 # Larger t for early aggressive exploration\n max_lp_iterations = 50 # Limit LP iterations per step\n\n while iteration < max_iterations and time.time() < deadline:\n iteration += 1\n print(f\"LP Iteration {iteration}\")\n\n # Compute convolution and max_b\n b_sequence = np.convolve(current_seq, current_seq)\n M = np.max(b_sequence)\n\n # Choose top K positions for constraints\n top_k_positions = np.argsort(b_sequence)[-K:]\n constraints = []\n\n # Add constraints for top K positions\n for x in top_k_positions:\n coeff = np.zeros(n)\n for j in range(n):\n if 0 <= x - j < n:\n coeff[j] = current_seq[x - j]\n constraints.append(cp.sum(coeff * cp.Variable(n, nonneg=True)) <= M)\n\n # Set up LP to maximize sum(b) under constraints\n b = cp.Variable(n, nonneg=True)\n objective = cp.Maximize(cp.sum(b))\n prob = cp.Problem(objective, constraints)\n\n # Solve the LP with ECOS for efficiency\n try:\n prob.solve(solver=cp.ECOS, verbose=False, maxiters=max_lp_iterations)\n if prob.status != cp.OPTIMAL:\n # Fallback: use current sequence if LP is not solvable\n b_opt = current_seq\n else:\n b_opt = b.value\n sum_b_opt = np.sum(b_opt)\n if sum_b_opt < 0.01:\n scaling_factor = 0.01 / sum_b_opt\n b_opt = b_opt * scaling_factor\n except:\n print(\"LP error, using current sequence\")\n b_opt = current_seq\n\n # Blend with current sequence\n t = t_initial * np.exp(-iteration / 20)\n new_seq = (1 - t) * current_seq + t * b_opt\n new_seq = np.clip(new_seq, 0, 1000.0)\n\n # Add small random perturbation to escape local minima\n noise = np.random.normal(0, 0.01, size=new_seq.shape)\n new_seq += noise\n new_seq = np.clip(new_seq, 0, 1000.0)\n\n # Evaluate new sequence\n new_score = evaluate_sequence(new_seq.tolist())\n if new_score < best_score:\n best_score = new_score\n best_seq = new_seq.copy()\n print(f\"New best score: {best_score}\")\n\n # Update current sequence for next iteration\n current_seq = new_seq\n\n # Ensure final sequence is valid\n best_seq = np.clip(best_seq, 0, 1000.0)\n return [float(x) for x in best_seq.tolist()]\n```",
"env/all/time/policy": 284.8152926629409,
"env/all/time/policy/min": 130.07353901863098,
"env/all/time/policy/max": 387.4623713493347,
"env/all/time/env_step": 821.9814522163942,
"env/all/time/env_step/min": 0.8498601913452148,
"env/all/time/env_step/max": 2194.5426433086395,
"env/all/time/reward_compute": 4.7031790018081665e-07,
"env/all/time/reward_compute/min": 1.825392246246338e-07,
"env/all/time/reward_compute/max": 1.2293457984924316e-06,
"env/all/by_group/frac_mixed": 1.0,
"env/all/by_group/frac_all_good": 0.0,
"env/all/by_group/frac_all_bad": 0.0,
"advantage/mean": 0.0054065459407866,
"advantage/min": -1.0,
"advantage/max": 1.2444167137145996,
"time/assemble_training_data": 6.092671632766724,
"time/kl_vs_base": 97.761150598526,
"kl_policy_base": 0.0006690432201139629,
"time/train": 544.1149432659149,
"time/save_checkpoint": 32.40869665145874,
"time/total": 3244.2333002090454
}
[2026-07-09T06:24:18+00:00] job=1812628 node=node-12 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T08:30:32+00:00] job=1813127 node=node-22 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T13:07:08+00:00] job=1813128 node=node-6 ngpu=6 ntrain=2 replicas=4 flash_attn=yes
[2026-07-11T13:13:37+00:00] job=1824165 node=node-6 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-12T04:58:24+00:00] job=1827204 node=node-4 ngpu=6 ntrain=2 replicas=4 flash_attn=yes