prithivMLmods commited on
Commit
52f5a95
·
verified ·
1 Parent(s): 71575b1

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +21 -0
README.md CHANGED
@@ -1,7 +1,28 @@
1
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2
  base_model:
3
  - harness-hub/Qwen3.8-27B-opencode-adapted
 
 
4
  ---
 
5
  # **Qwen3.8-27B-opencode-adapted-bf16**
6
 
7
  > **Qwen3.8-27B-opencode-adapted-bf16** is a standalone BF16 checkpoint created by merging the **harness-hub/Qwen3.8-27B-opencode-adapted** LoRA adapter into `Qwen/Qwen3.8-27B`, enabling deployment on any inference engine without LoRA support. The adapter (r = 32, alpha = 64, targeting the q, k, v, o, gate, up, and down projections) was trained with AsyncGRPO reinforcement learning in Harbor/OpenEnv environments to serve as the backbone for the OpenCode coding agent on software-engineering and terminal tasks, with thinking mode turned off as in training. The merge applies the standard `W + (alpha/r)·B@A` update per tensor in fp32 before casting back to BF16, preserving the base shard layout, tensor names, and all MTP tensors. According to the adapter's model card, under OpenCode it raises SWE-bench Pro (hard 51) from 35.3% to 51.0% and Terminal-Bench 2.1 from 46.4% to 52.4%, while requiring a quarter to half as many turns and generating 6–8× fewer tokens per task. These figures were measured using the adapter on the BF16 base model and have not been re-evaluated on this merged checkpoint.
 
1
  ---
2
+ license: apache-2.0
3
+ datasets:
4
+ - harness-hub/scale800-harbor
5
+ - harness-hub/harness-adapters
6
+ - harness-hub/termgrade-train-harbor
7
+ - ai-and/termgrade-environments
8
+ tags:
9
+ - mtp
10
+ - text-generation-inference
11
+ - trl
12
+ - grpo
13
+ - async-grpo
14
+ - lora
15
+ - agents
16
+ - harbor
17
+ - openenv
18
+ - opencode
19
+ pipeline_tag: image-text-to-text
20
  base_model:
21
  - harness-hub/Qwen3.8-27B-opencode-adapted
22
+ language:
23
+ - en
24
  ---
25
+
26
  # **Qwen3.8-27B-opencode-adapted-bf16**
27
 
28
  > **Qwen3.8-27B-opencode-adapted-bf16** is a standalone BF16 checkpoint created by merging the **harness-hub/Qwen3.8-27B-opencode-adapted** LoRA adapter into `Qwen/Qwen3.8-27B`, enabling deployment on any inference engine without LoRA support. The adapter (r = 32, alpha = 64, targeting the q, k, v, o, gate, up, and down projections) was trained with AsyncGRPO reinforcement learning in Harbor/OpenEnv environments to serve as the backbone for the OpenCode coding agent on software-engineering and terminal tasks, with thinking mode turned off as in training. The merge applies the standard `W + (alpha/r)·B@A` update per tensor in fp32 before casting back to BF16, preserving the base shard layout, tensor names, and all MTP tensors. According to the adapter's model card, under OpenCode it raises SWE-bench Pro (hard 51) from 35.3% to 51.0% and Terminal-Bench 2.1 from 46.4% to 52.4%, while requiring a quarter to half as many turns and generating 6–8× fewer tokens per task. These figures were measured using the adapter on the BF16 base model and have not been re-evaluated on this merged checkpoint.