Downloads · 30 days
10
23% of all-time downloads
forseasons/fgpo-qwen2.5-1.5b-search-coldstart
fgpo-qwen2.5-1.5b-search-coldstart is a text generation model from forseasons. Use it when you need the model to write or continue text. It is set up for transformers. The card lists the license as apache-2.0.
This private repository contains the merged cold-start SFT checkpoint used by the Failure-Grounded Policy Optimization (FGPO) experiments. It is not the final RL-trained FGPO model and should not be cited as evidence…
Downloads · 30 days
10
23% of all-time downloads
All-time downloads
43
Public
Parameters
1.5B
3.1 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors3.1 GB · 99%
From the Hugging Face model README
This private repository contains the merged cold-start SFT checkpoint used by the Failure-Grounded Policy Optimization (FGPO) experiments. It is not the final RL-trained FGPO model and should not be cited as evidence that FGPO outperforms a baseline.
Qwen/Qwen2.5-1.5B-Instruct21e99618386fd24084890cacaf25af645382f322df1c7b93fb39070c52fcc1c182c8b06ebab36962c21160bd8827ea45aa2aa9822ee643638528384db6928637c82d5af16eaeb2d15ed83f34c7e14019a8a1f369fgpo_model_manifest.json records the byte size and SHA-256 digest of every original
checkpoint file. README.md and LICENSE are Hub metadata added after the frozen tree
was produced and are therefore not part of that nine-file tree digest.
For reliable Hub transfer, the original consolidated model.safetensors is published as
12 native Transformers safetensors shards plus model.safetensors.index.json. No tensor
was quantized or converted: all 338 tensors were checked with exact torch.equal against
the consolidated source. fgpo_hf_packaging_manifest.json binds the source digest to each
shard and records the exact-equality verification.
The checkpoint follows the Search-R1-style multi-turn protocol:
<search>query</search>
<information>retrieved text</information>
<answer>final answer</answer>
It is intended for controlled research on small search agents, cold-start evaluation, and reproducibility checks. A retriever and the matching prompt/verifier contract are required for meaningful agent evaluation.
The training examples were produced from execution-verified teacher traces over standard QA resources and then filtered by provenance and leakage gates. Raw benchmark questions, answers, Wikipedia passages, retrieval indices, and teacher traces are intentionally not redistributed in this repository. Users must obtain upstream resources under their own licenses and terms.
The base checkpoint and this derivative are distributed under the Apache License 2.0.
See LICENSE. Users remain responsible for the licenses and terms of any external
datasets, retrievers, and corpora they combine with the model.