LFM2.5-2.6B-multiharness-SFT
A full fine-tune of LiquidAI/LFM2.5-2.6B for agentic data-analysis tasks, trained with supervised fine-tuning (TRL SFT) using OpenCode, Claude Code, Codex, Mini-SWE-Agent. This release is step 4,484 on main, epoch 2, scoring 43.1% pass@1 across four evaluation harnesses.
Article · Collection · Evaluation tasks · Training dashboard
Training and evaluation curves
Training curves use a trailing 50-step mean (at least 10 observations). Evaluation markers show measured checkpoints; hollow markers and dotted segments indicate incomplete coverage. Faint background lines show the full recorded trajectory; colored lines reveal the measured checkpoints. The first frame shows the completed chart before replaying, so previews also contain the full curves. Tool-call savings compare tasks solved by both the base model and checkpoint, with a different matched cohort at each checkpoint. SFT evaluations were recorded at the end of each epoch only. The animation stops at this revision's step 4,484.
Static chart · Plotted data · Interactive article
Training
TRL SFT on FineEnvs/SmolDataEnvs-multiharness-sft, combining OpenCode, Claude Code, Codex and Mini-SWE-Agent examples without harness balancing. The data contains successful Qwen3.8-27B teacher rollouts collected through OpenEnv × Harbor in Daytona sandboxes. Up to three teacher attempts were allowed per task/harness; the first successful trace was retained.
This run uses 17,929 assistant-turn examples from 3,189 successful rollouts covering 888 distinct tasks. Loss is applied only to the next assistant response; user messages, tool outputs and prior history are context. Responses are tokenized for LFM, not trained against teacher token IDs or logprobs.
| Setting | Value |
|---|---|
| Epochs / checkpoint | 2 / 4,484 |
| Learning rate / schedule | 3e-6 / constant |
| Effective batch size | 8 |
| Optimizer / precision | paged AdamW 8-bit / bfloat16 |
| Training harnesses | OpenCode, Claude Code, Codex, Mini-SWE-Agent |
| Objective | Completion-only supervised fine-tuning |
Evaluation
250 fixed SmolDataEnvs test tasks (33 easy, 118 medium, 99 hard), each evaluated under four harnesses: 1,000 graded task/harness cells. Pass@1 uses the first graded attempt per cell; infrastructure retries do not turn it into pass@k. All cells below are graded.
| Harness | Correct / evaluated | Pass@1 |
|---|---|---|
| OpenCode | 95 / 250 | 38.0% |
| Claude Code | 106 / 250 | 42.4% |
| Codex | 117 / 250 | 46.8% |
| Mini-SWE-Agent | 113 / 250 | 45.2% |
| Overall | 431 / 1,000 | 43.1% |
Harness versions: OpenCode 1.18.31, Claude Code 2.1.270, Codex 0.154.0, Mini-SWE-Agent 2.4.6. Full scores, including difficulty breakdowns, are in eval_results.json.
These are single-run results on a specific task set and harness versions. Harness mix, training exposure and compute differ across runs; the scores do not isolate a causal effect of the harness or objective.
Load the checkpoint
The repository contains full saved model weights and the saved tokenizer/chat template, not a LoRA adapter. Training used Transformers 5.14.1; use a compatible Transformers release.
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
model_id = "FineEnvs/LFM2.5-2.6B-multiharness-SFT"
revision = "main"
tokenizer = AutoTokenizer.from_pretrained(model_id, revision=revision)
model = AutoModelForCausalLM.from_pretrained(
model_id, revision=revision, dtype=torch.bfloat16, device_map="auto"
)
Reproducing the task scores requires the agent harness and tools described in the article; a plain chat prompt is not the same evaluation.
Related releases
- LFM2.5-2.6B-multiharness-RL
- LFM2.5-2.6B-opencode-RL
- LFM2.5-2.6B-multiharness-SFT
- LFM2.5-2.6B-opencode-SFT
- Qwen3.5-2B-multiharness-RL
- Qwen3.5-2B-opencode-RL
- Qwen3.5-2B-opencode-standalone-RL
All models, datasets, environments and the article are linked in the multi-harness RL collection.
License and provenance
Derived from LiquidAI/LFM2.5-2.6B under the LFM Open License v1.0. The base model's license is included unchanged in LICENSE. FineEnvs modified the weights by supervised fine-tuning; this is not an official Liquid AI release. See NOTICE and release_manifest.json for modification notices, the pinned base revision, checkpoint identity and file checksums. Optimizer, scheduler, RNG and trainer state are excluded.
Citation
For the experiment, methodology and interpretation, cite the main article:
@misc{kolavi2026multiharnessrl,
author = {Adithya S Kolavi},
title = {The ultimate guide to multi-harness RL},
year = {2026},
url = {https://huggingface.co/spaces/AdithyaSK/multi-harness-rl}
}