Downloads · 30 days
0
just1nseo/qwen3-4b-if-rlvr-subset4k-ep3
qwen3-4b-if-rlvr-subset4k-ep3 is a text generation model from just1nseo. Use it when you need the model to write or continue text. It is set up for transformers.
BF16 Transformers exports from the six A-series and four requested B-series IF-RLVR reward ablations. All arms use 3 epochs, batch size 256, 8 rollouts per prompt, temperature 1.0, top-p 0.95, top-k 20, 2,048 prompt t…
Downloads · 30 days
0
Access
Public
Updated Aug 17, 2026
Repo size
338 GB
Likes
0
Public
Click a slice to open those files.
.safetensors338 GB · 100%
From the Hugging Face model README
BF16 Transformers exports from the six A-series and four requested B-series IF-RLVR reward ablations. All arms use 3 epochs, batch size 256, 8 rollouts per prompt, temperature 1.0, top-p 0.95, top-k 20, 2,048 prompt tokens, and up to 8,192 response tokens.
Each 4,096-row experiment contains checkpoints at global_step_16,
global_step_32, and global_step_48. The 3,612-row B5 experiment contains
global_step_14, global_step_28, and global_step_42. Remote verification
manifests are stored under _manifests/.
| Order | Experiment directory | Reward arm |
|---|---|---|
| 1 | qwen3_4b_sub4k_a0_constraint_only_b256_ep3 | Constraint-only |
| 2 | qwen3_4b_sub4k_a3_baseline_floorzero_flipszeroed_b256_ep3 | Strict anchor baseline |
| 3 | qwen3_4b_sub4k_a2_softfloor_keepif_b256_ep3 | Soft floor |
| 4 | qwen3_4b_sub4k_a4_floorpenalty01_b256_ep3 | Floor penalty −0.1 |
| 5 | qwen3_4b_sub4k_a5_nofloor_upperonly_b256_ep3 | No lower floor |
| 6 | qwen3_4b_sub4k_a1_flipabstain_floorzero_b256_ep3 | Flip-abstain |
| 7 | qwen3_4b_sub4k_b5_dropflips_strict_b256_ep3 | Remove all 484 original flipped rows |
| 8 | qwen3_4b_sub4k_b6_widen10_mnl_strict_b256_ep3 | Mean-NLL bounds widened 10%, strict residual flips |
| 9 | qwen3_4b_sub4k_b7_widen10_mnl_flipabstain_b256_ep3 | Same 10% cache, residual flips abstain |
| 10 | qwen3_4b_sub4k_b8_widen20_mnl_strict_b256_ep3 | Mean-NLL bounds widened 20%, strict residual flips |
The following arms continue from their verified BF16 global_step_48 export
through cumulative steps 64 and 80 (epochs 4 and 5). The trainable actor is
warm-started from step 48, the frozen KL reference remains Qwen/Qwen3-4B, and
AdamW is freshly initialized at the same constant learning rate of 1e-6.
Native optimizer state is retained for automatic resume after step 64.
| Order | Experiment directory | Source arm |
|---|---|---|
| 1 | qwen3_4b_sub4k_a1_flipabstain_floorzero_b256_ep5_from_ep3 | A1 |
| 2 | qwen3_4b_sub4k_a2_softfloor_keepif_b256_ep5_from_ep3 | A2 |
| 3 | qwen3_4b_sub4k_a3_baseline_floorzero_flipszeroed_b256_ep5_from_ep3 | A3 |
| 4 | qwen3_4b_sub4k_a4_floorpenalty01_b256_ep5_from_ep3 | A4 |
| 5 | qwen3_4b_sub4k_b6_widen10_mnl_strict_b256_ep5_from_ep3 | B6 |
| 6 | qwen3_4b_sub4k_b7_widen10_mnl_flipabstain_b256_ep5_from_ep3 | B7 |
The uploader validates the local model configuration, tokenizer, safetensor index, tensor count and dtypes, then verifies the complete remote file set and every file size before marking a checkpoint complete.