You can deploy this model and start earning money today!

yujackein/
onereason-8b-lora-rl10-r3replay-interp075-kvo-r64a64

Name: onereason-8b-lora-rl10-r3replay-interp075-kvo-r64a64
Availability: InStock

peft

safetensors

recommendation

generative-recommendation

Views

No views yet

OneReason-8B RL10 + R3 Replay KVO Transfer LoRA

This is a platform-deployable rank-64/alpha-64 LoRA for the untouched OpenOneRec/OneReason-8B-pretrain-competition base. Unlike the earlier epoch-2 residual-only package, this repository includes the complete platform-compatible RL10 parent update in the same adapter.

Construction

The adapter is the exact sum of two rank-32/alpha-32 LoRA branches:

the platform-scored RL10 parent associated with total score 1.2402;
the epoch-2 R3 replay step20/25 interpolation-0.75 residual, where K/V/O factors use the interpolation and Q/MLP factors remain at step20.

Both branches have alpha / rank = 1. Their A matrices are concatenated by rows and their B matrices by columns, yielding

B64 @ A64 = B_RL10 @ A_RL10 + B_residual @ A_residual.

This identity is exact for the persisted LoRA branches. The artifact is not equivalent to epoch2 full SFT + residual: the epoch-2 dense base is intentionally absent so this adapter can load directly on the official base.

SHA-256 provenance:

RL10 parent: adcd1b777ef7f9868cbe12db956f3adbc2eaa342ad747033a603156eeec9f676
transferred residual: efc2152249a3958a7aa854105743c68c33428f706c54b5ba4b298df262f77e5f
composed adapter: 1a0707f6b0138136ba4ae9ac77a35e124d06483dc7c9db5d4121984e2293a867

Local screen versus RL10

All screens used local deterministic proxies without an external LLM Judge.

R0 balanced 400: char unigram F1 +0.001292; bigram F1 +0.000316; no empty, thinking, format, SID, or route leakage.
R0 reference-caption NLL: -0.020889 (lower is better), with 362/400 rows improving.
R2 128-row user proxy: overall +0.001166; JSON validity unchanged at 1.0.
R3 256-row recommendation proxy: full target log-probability +0.032605, restricted hierarchy log-probability -0.010786, and restricted s_c probability -0.014554.

The R3 evidence is mixed, so this is an exploratory platform candidate rather than a locally certified improvement.

Evaluation status

Official platform evaluation: pending.

yujackein/onereason-8b-lora-rl10-r3replay-interp075-kvo-r64a64

OneReason-8B RL10 + R3 Replay KVO Transfer LoRA

Construction

Local screen versus RL10

Evaluation status

yujackein/
onereason-8b-lora-rl10-r3replay-interp075-kvo-r64a64