ANLP mathematical reasoning: adapter archive

Preservation of 119 distinct adapters from 157 surviving source copies. Includes original GRPO specialists and mixed model, SFT warm-ups, matched SOLVE controls, and surviving Phase-2 checkpoints. OPD-trained adapters are excluded at the user's request. No base weights, optimizer states, credentials, or dataset examples are included.

Loading

from transformers import Qwen3_5ForConditionalGeneration
from peft import PeftModel
base = Qwen3_5ForConditionalGeneration.from_pretrained("Qwen/Qwen3.5-2B", revision="15852e8c16360a2fea060d615a32b45270f8a8fc", dtype="bfloat16")
model = PeftModel.from_pretrained(base, "ransom32/anlp-reasoning-transfer-adapters", subfolder="adapters/phase1-grpo/solve66")

Keep the original non-thinking template and task prompts when reproducing the study. See archive-manifest.json for checksums, source aliases and any failures. Adapter configs use public base IDs; their tensors are unmodified. Several historical intermediate checkpoints have been removed from scratch and are not recoverable from git.

Adapter	Size (MB)
phase1-grpo/judge26	250.1
phase1-grpo/mixed75	250.1
phase1-grpo/recon104	250.1
phase1-grpo/solve66	250.1
mix-control/ctrl-lr6.7e-6-s0/step75	250.1
mix-control/ctrl-lr6.7e-6-s1/step75	250.1
mix-control/ctrl-s0/step75	250.1
mix-control/ctrl-lr1e-5-s1/step75	250.1
mix-control/ctrl-lr1e-5-s0/step75	250.1
mix-control/ctrl-s1/step75	250.1
phase1-sft/SJ-sft470	250.1
phase1-sft/SS-sft470	250.1
phase1-sft/SS2-sft470-nofilter	250.1
phase1-sft/sft470-adapter	250.1
mix-control/ctrl-lr6.7e-6-s0/step32	250.1
mix-control/ctrl-lr6.7e-6-s0/step40	250.1
mix-control/ctrl-lr6.7e-6-s1/step32	250.1
mix-control/ctrl-lr6.7e-6-s1/step40	250.1
failure-driven/armA-s1/step32	250.1
failure-driven/armA-s1/step40	250.1
failure-driven/armR-prefix/step32	250.1
failure-driven/armR-prefix/step40	250.1
mix-control/ctrl-s0/step32	250.1
mix-control/ctrl-s0/step40	250.1
failure-driven/armB2/step32	250.1
failure-driven/armB2/step40	250.1
failure-driven/armBown/step32	250.1
failure-driven/armBown/step40	250.1
failure-driven/armD2/step32	250.1
failure-driven/armD2/step40	250.1
failure-driven/armE/step32	250.1
failure-driven/armE/step40	250.1
failure-driven/armH/step32	250.1
failure-driven/armH/step40	250.1
failure-driven/armA/step32	250.1
failure-driven/armA/step40	250.1
mix-control/ctrl-lr1e-5-s1/step32	250.1
mix-control/ctrl-lr1e-5-s1/step40	250.1
failure-driven/armD2-s1/step32	250.1
failure-driven/armD2-s1/step40	250.1
mix-control/ctrl-lr1e-5-s0/step32	250.1
mix-control/ctrl-lr1e-5-s0/step40	250.1
mix-control/ctrl-s1/step32	250.1
mix-control/ctrl-s1/step40	250.1
failure-driven/r1/step8	250.1
mix-control/ctrl-lr6.7e-6-s0/best-step43	250.1
mix-control/ctrl-lr6.7e-6-s0/best-step74	250.1
mix-control/ctrl-lr6.7e-6-s0/step16	250.1
mix-control/ctrl-lr6.7e-6-s0/step24	250.1
mix-control/ctrl-lr6.7e-6-s0/step48	250.1
mix-control/ctrl-lr6.7e-6-s0/step56	250.1
mix-control/ctrl-lr6.7e-6-s0/step64	250.1
mix-control/ctrl-lr6.7e-6-s0/step72	250.1
mix-control/ctrl-lr6.7e-6-s0/step8	250.1
mix-control/ctrl-lr6.7e-6-s1/best-step44	250.1
mix-control/ctrl-lr6.7e-6-s1/best-step74	250.1
mix-control/ctrl-lr6.7e-6-s1/step16	250.1
mix-control/ctrl-lr6.7e-6-s1/step24	250.1
mix-control/ctrl-lr6.7e-6-s1/step48	250.1
mix-control/ctrl-lr6.7e-6-s1/step56	250.1
mix-control/ctrl-lr6.7e-6-s1/step64	250.1
mix-control/ctrl-lr6.7e-6-s1/step72	250.1
mix-control/ctrl-lr6.7e-6-s1/step8	250.1
failure-driven/armA-s1/step16	250.1
failure-driven/armA-s1/step24	250.1
failure-driven/armR-prefix/step16	250.1
failure-driven/armR-prefix/step24	250.1
mix-control/ctrl-s0/best-step25	250.1
mix-control/ctrl-s0/best-step44	250.1
mix-control/ctrl-s0/step16	250.1
mix-control/ctrl-s0/step24	250.1
mix-control/ctrl-s0/step48	250.1
mix-control/ctrl-s0/step56	250.1
mix-control/ctrl-s0/step64	250.1
mix-control/ctrl-s0/step72	250.1
mix-control/ctrl-s0/step8	250.1
failure-driven/armB/step16	250.1
failure-driven/armB/step24	250.1
failure-driven/armB2/step16	250.1
failure-driven/armB2/step24	250.1
failure-driven/armBown/step16	250.1
failure-driven/armBown/step24	250.1
failure-driven/armD/step16	250.1
failure-driven/armD/step24	250.1
failure-driven/armD2/step16	250.1
failure-driven/armD2/step24	250.1
failure-driven/armE/step16	250.1
failure-driven/armE/step24	250.1
failure-driven/armH/step16	250.1
failure-driven/armH/step24	250.1
mix-control/ctrl-lr1e-5-s1/best-step42	250.1
mix-control/ctrl-lr1e-5-s1/best-step74	250.1
mix-control/ctrl-lr1e-5-s1/step16	250.1
mix-control/ctrl-lr1e-5-s1/step24	250.1
mix-control/ctrl-lr1e-5-s1/step48	250.1
mix-control/ctrl-lr1e-5-s1/step56	250.1
mix-control/ctrl-lr1e-5-s1/step64	250.1
mix-control/ctrl-lr1e-5-s1/step72	250.1
mix-control/ctrl-lr1e-5-s1/step8	250.1
failure-driven/armD2-s1/step16	250.1
failure-driven/armD2-s1/step24	250.1
mix-control/ctrl-lr1e-5-s0/best-step42	250.1
mix-control/ctrl-lr1e-5-s0/best-step44	250.1
mix-control/ctrl-lr1e-5-s0/step16	250.1
mix-control/ctrl-lr1e-5-s0/step24	250.1
mix-control/ctrl-lr1e-5-s0/step48	250.1
mix-control/ctrl-lr1e-5-s0/step56	250.1
mix-control/ctrl-lr1e-5-s0/step64	250.1
mix-control/ctrl-lr1e-5-s0/step72	250.1
mix-control/ctrl-lr1e-5-s0/step8	250.1
mix-control/ctrl-s1/best-step42	250.1
mix-control/ctrl-s1/best-step74	250.1
mix-control/ctrl-s1/step16	250.1
mix-control/ctrl-s1/step24	250.1
mix-control/ctrl-s1/step48	250.1
mix-control/ctrl-s1/step56	250.1
mix-control/ctrl-s1/step64	250.1
mix-control/ctrl-s1/step72	250.1
mix-control/ctrl-s1/step8	250.1

Downloads last month: -

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for ransom32/anlp-reasoning-transfer-adapters

Base model

Qwen/Qwen3.5-2B-Base

Finetuned

Qwen/Qwen3.5-2B

Adapter

(279)

this model